diff --git a/test_config_123.yaml b/test_config_123.yaml index f08e0ca5703..459685f50ec 100644 --- a/test_config_123.yaml +++ b/test_config_123.yaml @@ -1,26 +1,3605 @@ #################################################################### -### Simplified LiteLLM Proxy Configuration +### EMEA PROD LiteLLM Proxy Configuration #################################################################### +# Global API base for all model deployments (YAML anchor) +x-api-base: &api_base http://localhost:8090 + general_settings: - master_key: sk-1234 - disable_prisma_schema_update: true - database_url: + master_key: REDACTED_MASTER_KEY # set via env or replace for local use + background_health_checks: false + disable_spend_updates: true + health_check_interval: 300 # frequency of background health checks every 5 minutes (kept for reference) + health_check_details: false # Disable health check details for comprehensive testing + + # NOTE: MANAGED BY CCoE. Set via env or replace with real URL. + database_url: postgresql://REDACTED_USER:REDACTED_PASSWORD@REDACTED_HOST/neondb?sslmode=require + disable_prisma_schema_update: true # Disable Prisma migrations on startup + + # TODO: The next two settings are in the best practices section, but they are not enabled by default. + # TODO: Discuss with Javier first, before you enable this + proxy_batch_write_at: 60 # Batch write spend updates every 60s + # database_connection_pool_limit: 10 # limit the number of database connections to = MAX Number of DB Connections/Number of instances of litellm proxy (Around 10-20 is good number) + + + use_redis_transaction_buffer: true # avoid deadlocks, see https://docs.litellm.ai/docs/proxy/db_deadlocks#stage-2-a-single-instance-flushes-the-redis-queue-to-the-db + + ## TODO: We might want to enable this in the future, but for now we disable it + store_model_in_db: true # Disable DB storage to use file-based config only + supported_db_objects: [] + # MS Teams Weboook + # see also: https://docs.litellm.ai/docs/proxy/alerting#ms-teams-webhooks + alerting: ["slack"] + alerting_threshold: 300 # sends alerts if requests hang for 5min+ and responses take 5min+ + # See also: https://docs.litellm.ai/docs/proxy/alerting#all-possible-alert-types + alert_types: [ + "llm_exceptions", + "llm_too_slow", + "llm_requests_hanging", + "budget_alerts", + "spend_reports", + "db_exceptions", + "daily_reports", + "cooldown_deployment", + "new_model_added", + "outage_alerts", + "new_virtual_key_created", + "failed_tracking_spend", + "db_deadlocks", # Critical for database issues + "memory_usage_high" # Critical for memory leak detection + ] ################################# -### LiteLLM Settings +### ADVANCED PROMETHEUS CONFIGURATION +### Documentation: https://docs.litellm.ai/docs/proxy/prometheus ################################# litellm_settings: - callbacks: ["prometheus"] - service_callback: ["prometheus_system"] + # Cache + cache: true + # NOTE: MANAGED BY CCoE. Set host/password via env or replace for local use. + cache_params: + type: redis + host: REDACTED_REDIS_HOST + port: 17950 + username: default + password: REDACTED_REDIS_PASSWORD + + # ============================================================================= + # PROMETHEUS CORE CONFIGURATION + # ============================================================================= + + # # Enable Prometheus metrics collection, Langfuse integration, and OpenTelemetry + # # This exposes /metrics endpoint for Prometheus to scrape + # # Using standard langfuse callback for now - langfuse_otel has compatibility issues + # # Global Langfuse enabled for fallback when no team-specific settings + # # OpenTelemetry enabled for enterprise observability platform integration + # callbacks: ["prometheus", "langfuse"] + + + + # # ============================================================================= + # # SYSTEM HEALTH & INFRASTRUCTURE MONITORING + # # ============================================================================= + + # # Enable system-level health monitoring for adjacent services + # # Monitors Redis latency, database performance, and LiteLLM self-performance + # # Metrics emitted: litellm_redis_latency, litellm_redis_fails, litellm_self_latency + # service_callback: ["prometheus_system"] + + # ============================================================================= + # SECURITY & ACCESS CONTROL + # ============================================================================= + + # Require authentication for /metrics endpoint access + # This prevents unauthorized access to sensitive usage and performance data + # Set to true in production environments + require_auth_for_metrics_endpoint: false + + # ============================================================================= + # PERFORMANCE & DEBUGGING SETTINGS + # ============================================================================= + + # Keep API key info visible for detailed tracking + redact_user_api_key_info: false + + # Enable comprehensive audit logging for compliance and debugging + store_audit_logs: true + + # Use structured JSON logs for better integration with log aggregation systems + json_logs: true # Better for log aggregation + + # ============================================================================= + # DATABASE OPTIMIZATION (ADVANCED MONITORING FOCUS) + # ============================================================================= + + # Disable detailed spend/error logs to reduce DB load while keeping Prometheus metrics + # This optimizes for metrics-based monitoring rather than detailed log analysis + disable_spend_logs: True # Disable writing spend logs to DB - use Prometheus metrics instead + disable_error_logs: True # Disable writing error logs to DB - use Prometheus alerts instead + + # ============================================================================= + # SELF-SERVE KEY GENERATION SETTINGS + # Documentation: https://docs.litellm.ai/docs/proxy/self_serve + # ============================================================================= + + # Configure who can generate personal keys through the self-serve interface + # This enables controlled key generation for authorized users only + key_generation_settings: + team_key_generation: + allowed_team_member_roles: ["proxy_admin"] # Only proxy admins and team admins can generate team keys + personal_key_generation: + allowed_user_roles: ["proxy_admin"] # Only proxy admins can generate personal keys + + + # ============================================================================= + # General LANGFUSE Settings - Standard Integration + # ============================================================================= + + # Enable Langfuse tracking for all teams with default tags using standard protocol + # Standard langfuse callback provides reliable trace collection for now + # + # Documentation: https://docs.litellm.ai/docs/observability/langfuse_integration + langfuse_default_tags: ["cache_hit", "cache_key", "proxy_base_url", "user_api_key_alias", "user_api_key_user_id", "user_api_key_user_email", "user_api_key_team_alias", "semantic-similarity", "proxy_base_url"] + ################################# -### Models - One model only +### Router Settings - Fallbacks ################################# + +router_settings: + # Simple string mapping (alias -> model) + model_group_alias: { + "gpt-o1": "gpt-o1-20241217", + "gpt-4o-global": "gpt-4o-20241120-global", + "gpt-4": "gpt-4", + "gpt-35-turbo": "gpt-35-turbo", + "eu-gpt-o1": "gpt-o1-20241217", + "na-gpt-o1": "gpt-o1-20241217", + "eu-gpt-4o": "gpt-4o-20241120-global", + "na-gpt-4o": "gpt-4o-20241120-global", + "gpt-4o-load-balanced": "gpt-4o-global-balanced" + } + + enable_tag_filtering: True + routing_strategy: "simple-shuffle" # Best optimised for PROD environments + cooldown_time: 30 # Faster recovery than default + num_retries: 2 # Faster failure detection + + fallbacks: + - {"gpt-4o-20241120-global": ["gpt-4o-20241120-global"]} + - {"gpt-4": ["gpt-4"]} + - {"gpt-4o-global-balanced": ["gpt-4o-global-balanced"]} + - {"gpt-o1-20241217": ["gpt-o1-20241217"]} + - {"gpt-35-turbo": ["gpt-35-turbo"]} + - {"gpt-35-turbo-16k": ["gpt-35-turbo-16k"]} + - {"text-embedding-3-large": ["text-embedding-3-large"]} + - {"gpt-4o-realtime-preview": ["gpt-4o-realtime-preview"]} + +################################# +### Models - Azure OpenAI +################################# + model_list: - - model_name: "gpt-o1" + # TODO: Reconfigure the load balanced models with the new instances, as now we have more of them having this deployments names to be load-balanced + # Multi-Region - Load Balanced through routing strategy + - model_name: "gpt-o1-20241217" litellm_params: - model: openai/gpt-o1 - api_base: https://exampleopenaiendpoint-production-0ee2.up.railway.app/ + model: azure/gpt-o1-20241217 + base_model: "o1-2024-12-17" + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "sweden-central", "azure"] model_info: - context_length: 200000 \ No newline at end of file + base_model: "o1-2024-12-17" + access_groups: ["Default", "EMEA"] + context_length: 200000 + - model_name: "gpt-o1-20241217" + litellm_params: + model: azure/gpt-o1-20241217 + base_model: "o1-2024-12-17" + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us-2", "azure"] + model_info: + base_model: "o1-2024-12-17" + access_groups: ["Default", "NA"] + context_length: 200000 + - model_name: "gpt-4o-global" + litellm_params: + model: azure/gpt-4o-20241120-global + base_model: "gpt-4o" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us-2", "azure"] + model_info: + base_model: "gpt-4o" + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-4o-20241120-datazone" + litellm_params: + model: azure/gpt-4o-20241120-datazone + base_model: "gpt-4o" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us-2", "azure"] + model_info: + base_model: "gpt-4o" + access_groups: ["Default", "NA"] + context_length: 65536 + - model_name: "gpt-4o-global" + litellm_params: + model: azure/gpt-4o-20241120-global + base_model: "gpt-4o" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "sweden-central", "azure"] + model_info: + base_model: "gpt-4o" + access_groups: ["Default", "EMEA"] + context_length: 128000 + + # ============================================ + # Sweden Central (sc) - Chat Completion Models + # ============================================ + - model_name: "gpt-35-turbo" + litellm_params: + model: azure/gpt-35-turbo + base_model: "gpt-3.5-turbo" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "sweden-central", "azure"] + model_info: + base_model: "gpt-3.5-turbo" + access_groups: ["Default", "EMEA"] + context_length: 128000 + - model_name: "gpt-o3-20250416-global" + litellm_params: + model: azure/gpt-o3-20250416-global + base_model: "o3" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "sweden-central", "azure"] + model_info: + base_model: "o3" + access_groups: ["Default", "EMEA"] + context_length: 200000 + - model_name: "gpt-5.1-20251113-global" + litellm_params: + model: azure/gpt-5.1-20251113-global + base_model: "gpt-5.1" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "sweden-central", "azure"] + model_info: + base_model: "gpt-5.1" + access_groups: ["Default", "EMEA"] + context_length: 128000 + - model_name: "gpt-5.1-20251113-datazone" + litellm_params: + model: azure/gpt-5.1-20251113-datazone + base_model: "gpt-5.1" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "sweden-central", "azure"] + model_info: + base_model: "gpt-5.1" + access_groups: ["Default", "EMEA"] + context_length: 128000 + - model_name: "gpt-5.2-20251211-global" + litellm_params: + model: azure/gpt-5.2-20251211-global + base_model: "gpt-5.2" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "sweden-central", "azure"] + model_info: + base_model: "gpt-5.2" + access_groups: ["Default", "EMEA"] + context_length: 128000 + - model_name: "gpt-5-20250807-global" + litellm_params: + model: azure/gpt-5-20250807-global + base_model: "gpt-5" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "sweden-central", "azure"] + model_info: + base_model: "gpt-5" + access_groups: ["Default", "EMEA"] + context_length: 128000 + - model_name: "gpt-5-mini-20250807-global" + litellm_params: + model: azure/gpt-5-mini-20250807-global + base_model: "gpt-5-mini" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "sweden-central", "azure"] + model_info: + base_model: "gpt-5-mini" + access_groups: ["Default", "EMEA"] + context_length: 128000 + - model_name: "gpt-5-nano-20250807-global" + litellm_params: + model: azure/gpt-5-nano-20250807-global + base_model: "gpt-5-nano" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "sweden-central", "azure"] + model_info: + base_model: "gpt-5-nano" + access_groups: ["Default", "EMEA"] + context_length: 128000 + - model_name: "gpt-4.1-20250414-datazone" + litellm_params: + model: azure/gpt-4.1-20250414-datazone + base_model: "gpt-4" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "sweden-central", "azure"] + model_info: + base_model: "gpt-4" + access_groups: ["Default", "EMEA"] + context_length: 128000 + - model_name: "gpt-4.1-20250414-global" + litellm_params: + model: azure/gpt-4.1-20250414-global + base_model: "gpt-4" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "sweden-central", "azure"] + model_info: + base_model: "gpt-4" + access_groups: ["Default", "EMEA"] + context_length: 128000 + - model_name: "gpt-4.1-mini-20250414-global" + litellm_params: + model: azure/gpt-4.1-mini-20250414-global + base_model: "gpt-4o-mini" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "sweden-central", "azure"] + model_info: + base_model: "gpt-4o-mini" + access_groups: ["Default", "EMEA"] + context_length: 128000 + - model_name: "gpt-4.1-nano-20250414-datazone" + litellm_params: + model: azure/gpt-4.1-nano-20250414-datazone + base_model: "gpt-4o-mini" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "sweden-central", "azure"] + model_info: + base_model: "gpt-4o-mini" + access_groups: ["Default", "EMEA"] + context_length: 128000 + - model_name: "gpt-4.1-nano-20250414-global" + litellm_params: + model: azure/gpt-4.1-nano-20250414-global + base_model: "gpt-4o-mini" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "sweden-central", "azure"] + model_info: + base_model: "gpt-4o-mini" + access_groups: ["Default", "EMEA"] + context_length: 128000 + - model_name: "gpt-4o-20240513-global" + litellm_params: + model: azure/gpt-4o-20240513-global + base_model: "gpt-4o" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "sweden-central", "azure"] + model_info: + base_model: "gpt-4o" + access_groups: ["Default", "EMEA"] + context_length: 128000 + - model_name: "GPT-4o-20240513-standard" + litellm_params: + model: azure/GPT-4o-20240513-standard + base_model: "gpt-4o" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "sweden-central", "azure"] + model_info: + base_model: "gpt-4o" + access_groups: ["Default", "EMEA"] + context_length: 128000 + - model_name: "gpt-4o-20240806-datazone" + litellm_params: + model: azure/gpt-4o-20240806-datazone + base_model: "gpt-4o" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "sweden-central", "azure"] + model_info: + base_model: "gpt-4o" + access_groups: ["Default", "EMEA"] + context_length: 128000 + - model_name: "gpt-4o-20240806-standard" + litellm_params: + model: azure/gpt-4o-20240806-standard + base_model: "gpt-4o" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "sweden-central", "azure"] + model_info: + base_model: "gpt-4o" + access_groups: ["Default", "EMEA"] + context_length: 128000 + - model_name: "gpt-4o-20241120-global" + litellm_params: + model: azure/gpt-4o-20241120-global + base_model: "gpt-4o" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "sweden-central", "azure"] + model_info: + base_model: "gpt-4o" + access_groups: ["Default", "EMEA"] + context_length: 128000 + - model_name: "gpt-4o-20241120-standard" + litellm_params: + model: azure/gpt-4o-20241120-standard + base_model: "gpt-4o" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "sweden-central", "azure"] + model_info: + base_model: "gpt-4o" + access_groups: ["Default", "EMEA"] + context_length: 128000 + - model_name: "gpt-4o-mini-20240718" + litellm_params: + model: azure/gpt-4o-mini-20240718 + base_model: "gpt-4o-mini" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "sweden-central", "azure"] + model_info: + base_model: "gpt-4o-mini" + access_groups: ["Default", "EMEA"] + context_length: 128000 + - model_name: "gpt-4o-mini-20240718-datazone" + litellm_params: + model: azure/gpt-4o-mini-20240718-datazone + base_model: "gpt-4o-mini" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "sweden-central", "azure"] + model_info: + base_model: "gpt-4o-mini" + access_groups: ["Default", "EMEA"] + context_length: 128000 + - model_name: "gpt-4o-mini-20240718-standard" + litellm_params: + model: azure/gpt-4o-mini-20240718-standard + base_model: "gpt-4o-mini" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "sweden-central", "azure"] + model_info: + base_model: "gpt-4o-mini" + access_groups: ["Default", "EMEA"] + context_length: 128000 + - model_name: "gpt-o1-20241217" + litellm_params: + model: azure/gpt-o1-20241217 + base_model: "o1-2024-12-17" + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "sweden-central", "azure"] + model_info: + base_model: "o1-2024-12-17" + access_groups: ["Default", "EMEA"] + context_length: 200000 + - model_name: "GPT-o1-mini-20240912" + litellm_params: + model: azure/GPT-o1-mini-20240912 + base_model: "o1-mini" + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "sweden-central", "azure"] + model_info: + base_model: "o1-mini" + access_groups: ["Default", "EMEA"] + context_length: 200000 + - model_name: "gpt-o3-mini-20250131" + litellm_params: + model: azure/gpt-o3-mini-20250131 + base_model: "o3-mini" + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "sweden-central", "azure"] + model_info: + base_model: "o3-mini" + access_groups: ["Default", "EMEA"] + context_length: 200000 + - model_name: "gpt-o4-mini-20250416-global" + litellm_params: + model: azure/gpt-o4-mini-20250416-global + base_model: "o4-mini" + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "sweden-central", "azure"] + model_info: + base_model: "o4-mini" + access_groups: ["Default", "EMEA"] + context_length: 200000 + - model_name: "o1-20241217-datazone" + litellm_params: + model: azure/o1-20241217-datazone + base_model: "o1-2024-12-17" + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "sweden-central", "azure"] + model_info: + base_model: "o1-2024-12-17" + access_groups: ["Default", "EMEA"] + context_length: 200000 + - model_name: "o1-mini-20240912-standard" + litellm_params: + model: azure/o1-mini-20240912-standard + base_model: "o1-mini" + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "sweden-central", "azure"] + model_info: + base_model: "o1-mini" + access_groups: ["Default", "EMEA"] + context_length: 200000 + - model_name: "o3-20250416-datazone" + litellm_params: + model: azure/o3-20250416-datazone + base_model: "o3" + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "sweden-central", "azure"] + model_info: + base_model: "o3" + access_groups: ["Default", "EMEA"] + context_length: 200000 + - model_name: "o3-mini-20250131-datazone" + litellm_params: + model: azure/o3-mini-20250131-datazone + base_model: "o3-mini" + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "sweden-central", "azure"] + model_info: + base_model: "o3-mini" + access_groups: ["Default", "EMEA"] + context_length: 200000 + - model_name: "o4-mini-20250416-datazone" + litellm_params: + model: azure/o4-mini-20250416-datazone + base_model: "o4-mini" + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "sweden-central", "azure"] + model_info: + base_model: "o4-mini" + access_groups: ["Default", "EMEA"] + context_length: 200000 + + # ============================= + # Sweden Central - Audio Models + # ============================= + - model_name: "tts-1" + litellm_params: + model: azure/tts-001 + base_model: "tts-1" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "sweden-central", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + mode: "audio" + - model_name: "tts-1-hd" + litellm_params: + model: azure/tts-hd-001 + base_model: "tts-1-hd" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "sweden-central", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + mode: "audio" + - model_name: "whisper-1" + litellm_params: + model: azure/whisper + base_model: "whisper-1" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "sweden-central", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + mode: "audio" + - model_name: "gpt-4o-audio-preview-20241217" + litellm_params: + model: azure/gpt-4o-audio-preview-20241217 + base_model: "gpt-4o" + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "sweden-central", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + mode: "audio" + + # ======================================== + # Sweden Central - Image Generation Models + # ======================================== + - model_name: "dall-e-3" + litellm_params: + model: azure/dall-e-3 + base_model: "dall-e-3" + api_version: 2023-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "sweden-central", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + mode: "image" + + # ================================== + # Sweden Central - Embeddings Models + # ================================== + - model_name: "text-embedding-ada-002" + litellm_params: + model: azure/text-embedding-ada-002 + base_model: "text-embedding-ada-002" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "sweden-central", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + mode: "embedding" + - model_name: "text-embedding-3-large" + litellm_params: + model: azure/text-embedding-3-large + base_model: "text-embedding-3-large" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "sweden-central", "azure"] + model_info: + base_model: "text-embedding-3-large" + access_groups: ["Default", "EMEA"] + mode: "embedding" + - model_name: "text-embedding-3-small-1-global" + litellm_params: + model: azure/text-embedding-3-small-1-global + base_model: "text-embedding-3-small" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "sweden-central", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + mode: "embedding" + + # ======================================= + # Sweden Central (sc) - Transcribe Models + # ======================================= + - model_name: "gpt-4o-transcribe-20250320-global" + litellm_params: + model: azure/gpt-4o-transcribe-20250320-global + base_model: "gpt-4o-transcribe" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "sweden-central", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + mode: "audio_transcription" + - model_name: "gpt-4o-mini-transcribe-20250320-global" + litellm_params: + model: azure/gpt-4o-mini-transcribe-20250320-global + base_model: "gpt-4o-mini-transcribe" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "sweden-central", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + mode: "audio_transcription" + - model_name: "gpt-4o-transcribe-diarize-20251015-global" + litellm_params: + model: azure/gpt-4o-transcribe-diarize-20251015-global + base_model: "gpt-4o-transcribe-diarize" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "sweden-central", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + mode: "audio_transcription" + + # ========================================= + # East US 2 (eus2) - Chat Completion Models + # ========================================= + - model_name: "gpt-35-turbo" + litellm_params: + model: azure/gpt-35-turbo + base_model: "gpt-3.5-turbo" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us-2", "azure"] + model_info: + base_model: "gpt-3.5-turbo" + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-o3-20250416-global" + litellm_params: + model: azure/gpt-o3-20250416-global + base_model: "o3" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us-2", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 200000 + - model_name: "gpt-4.1-20250414-datazone" + litellm_params: + model: azure/gpt-4.1-20250414-datazone + base_model: "gpt-4" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us-2", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-4.1-20250414-global" + litellm_params: + model: azure/gpt-4.1-20250414-global + base_model: "gpt-4" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us-2", "azure"] + model_info: + base_model: "gpt-4" + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-4.1-mini-20250414-datazone" + litellm_params: + model: azure/gpt-4.1-mini-20250414-datazone + base_model: "gpt-4o-mini" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us-2", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-4.1-mini-20250414-global" + litellm_params: + model: azure/gpt-4.1-mini-20250414-global + base_model: "gpt-4o-mini" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us-2", "azure"] + model_info: + base_model: "gpt-4o-mini" + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-4.1-nano-20250414-datazone" + litellm_params: + model: azure/gpt-4.1-nano-20250414-datazone + base_model: "gpt-4o-mini" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us-2", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-4.1-nano-20250414-global" + litellm_params: + model: azure/gpt-4.1-nano-20250414-global + base_model: "gpt-4o-mini" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us-2", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-4o-2024-08-06-Standard" + litellm_params: + model: azure/gpt-4o-2024-08-06-Standard + base_model: "gpt-4o" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us-2", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-4o-20240513-global" + litellm_params: + model: azure/gpt-4o-20240513-global + base_model: "gpt-4o" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us-2", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "GPT-4o-20240513-standard" + litellm_params: + model: azure/GPT-4o-20240513-standard + base_model: "gpt-4o" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us-2", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-4o-20241120-global" + litellm_params: + model: azure/gpt-4o-20241120-global + base_model: "gpt-4o" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us-2", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-4o-20241120-standard" + litellm_params: + model: azure/gpt-4o-20241120-standard + base_model: "gpt-4o" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us-2", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-4o-mini-20240718" + litellm_params: + model: azure/gpt-4o-mini-20240718 + base_model: "gpt-4o-mini" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us-2", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-4o-mini-20240718-datazone" + litellm_params: + model: azure/gpt-4o-mini-20240718-datazone + base_model: "gpt-4o-mini" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us-2", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-4o-mini-20240718-standard" + litellm_params: + model: azure/gpt-4o-mini-20240718-standard + base_model: "gpt-4o-mini" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us-2", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-4o-mini-20240718-standard" + litellm_params: + model: azure/gpt-4o-mini-20240718-standard + base_model: "gpt-4o-mini" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us-2", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-5.1-20251113-global" + litellm_params: + model: azure/gpt-5.1-20251113-global + base_model: "gpt-5.1" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us-2", "azure"] + model_info: + base_model: "gpt-5.1" + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-5.1-20251113-datazone" + litellm_params: + model: azure/gpt-5.1-20251113-datazone + base_model: "gpt-5.1" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us-2", "azure"] + model_info: + base_model: "gpt-5.1" + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-5.2-20251211-global" + litellm_params: + model: azure/gpt-5.2-20251211-global + base_model: "gpt-5.2" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us-2", "azure"] + model_info: + base_model: "gpt-5.2" + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-5-20250807-global" + litellm_params: + model: azure/gpt-5-20250807-global + base_model: "gpt-5" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us-2", "azure"] + model_info: + base_model: "gpt-5" + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-5-mini-20250807-global" + litellm_params: + model: azure/gpt-5-mini-20250807-global + base_model: "gpt-5-mini" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us-2", "azure"] + model_info: + base_model: "gpt-5-mini" + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-5-nano-20250807-global" + litellm_params: + model: azure/gpt-5-nano-20250807-global + base_model: "gpt-5-nano" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us-2", "azure"] + model_info: + base_model: "gpt-5-nano" + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-o1-20241217" + litellm_params: + model: azure/gpt-o1-20241217 + base_model: "o1-2024-12-17" + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us-2", "azure"] + model_info: + base_model: "o1-2024-12-17" + access_groups: ["Default", "NA"] + context_length: 200000 + - model_name: "GPT-o1-mini-20240912" + litellm_params: + model: azure/GPT-o1-mini-20240912 + base_model: "o1-mini" + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us-2", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 200000 + - model_name: "gpt-o3-mini-20250131" + litellm_params: + model: azure/gpt-o3-mini-20250131 + base_model: "o3-mini" + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us-2", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 200000 + - model_name: "gpt-o4-mini-20250416-global" + litellm_params: + model: azure/gpt-o4-mini-20250416-global + base_model: "o4-mini" + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us-2", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 200000 + - model_name: "o1-20241217-datazone" + litellm_params: + model: azure/o1-20241217-datazone + base_model: "o1-2024-12-17" + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us-2", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 200000 + - model_name: "o1-mini-20240912-standard" + litellm_params: + model: azure/o1-mini-20240912-standard + base_model: "o1-mini" + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us-2", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 200000 + - model_name: "o3-20250416-datazone" + litellm_params: + model: azure/o3-20250416-datazone + base_model: "o3" + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us-2", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 200000 + - model_name: "o3-mini-20250131-datazone" + litellm_params: + model: azure/o3-mini-20250131-datazone + base_model: "o3-mini" + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us-2", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 200000 + - model_name: "o3-pro-20250610-global" + litellm_params: + model: azure/o3-pro-20250610-global + base_model: "o3" + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us-2", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 200000 + - model_name: "o4-mini-20250416-datazone" + litellm_params: + model: azure/o4-mini-20250416-datazone + base_model: "o4-mini" + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us-2", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 200000 + + # ============================= + # East US 2 - Embeddings Models + # ============================= + - model_name: "text-text-embedding-3-1-small-global" + litellm_params: + model: azure/text-embedding-3-1-small-global + base_model: "text-embedding-3-small" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us-2", "azure"] + model_info: + access_groups: ["Default", "NA"] + mode: "embedding" + - model_name: "text-embedding-3-large" + litellm_params: + model: azure/text-embedding-3-large + base_model: "text-embedding-3-large" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us-2", "azure"] + model_info: + base_model: "text-embedding-3-large" + access_groups: ["Default", "NA"] + mode: "embedding" + - model_name: "text-embedding-3-small" + litellm_params: + model: azure/text-embedding-3-small + base_model: "text-embedding-3-small" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us-2", "azure"] + model_info: + access_groups: ["Default", "NA"] + mode: "embedding" + - model_name: "text-embedding-ada-002" + litellm_params: + model: azure/text-embedding-ada-002 + base_model: "text-embedding-ada-002" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us-2", "azure"] + model_info: + access_groups: ["Default", "NA"] + mode: "embedding" + + # ==================================== + # East US 2 (eus2) - Transcribe Models + # ==================================== + - model_name: "gpt-4o-mini-transcribe-20250320-global" + litellm_params: + model: azure/gpt-4o-mini-transcribe-20250320-global + base_model: "gpt-4o-mini-transcribe" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us-2", "azure"] + model_info: + access_groups: ["Default", "NA"] + mode: "audio_transcription" + - model_name: "gpt-4o-transcribe-20250320-global" + litellm_params: + model: azure/gpt-4o-transcribe-20250320-global + base_model: "gpt-4o-transcribe" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us-2", "azure"] + model_info: + access_groups: ["Default", "NA"] + mode: "audio_transcription" + - model_name: "gpt-4o-transcribe-diarize-20251015-global" + litellm_params: + model: azure/gpt-4o-transcribe-diarize-20251015-global + base_model: "gpt-4o-transcribe-diarize" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us-2", "azure"] + model_info: + access_groups: ["Default", "NA"] + mode: "audio_transcription" + + # ======================== + # East US 2 - Audio Models + # ======================== + - model_name: "gpt-4o-audio-preview-20241217" + litellm_params: + model: azure/gpt-4o-audio-preview-20241217 + base_model: "gpt-4o" + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us-2", "azure"] + model_info: + access_groups: ["Default", "NA"] + mode: "audio" + - model_name: "gpt-4o-mini-tts-20240320-global" + litellm_params: + model: azure/gpt-4o-mini-tts-20240320-global + base_model: "tts-1" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us-2", "azure"] + model_info: + access_groups: ["Default", "NA"] + mode: "audio" + + # =================== + # Realtime API Models + # =================== + - model_name: "gpt-realtime-20250828-standard" + litellm_params: + model: azure/gpt-realtime-20250828-standard + base_model: "gpt-realtime-2025-08-28" + api_version: 2025-04-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "sweden-central", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + mode: "realtime" + context_length: 128000 + - model_name: "gpt-realtime-20250828-standard" + litellm_params: + model: azure/gpt-realtime-20250828-standard + base_model: "gpt-realtime-2025-08-28" + api_version: 2025-04-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us-2", "azure"] + model_info: + access_groups: ["Default", "NA"] + mode: "realtime" + context_length: 128000 + - model_name: "gpt-4o-realtime-preview-20241217-global" + litellm_params: + model: azure/gpt-4o-realtime-preview-20241217-global + base_model: "gpt-4o-realtime-preview-2024-12-17" + api_version: 2025-04-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "sweden-central", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + mode: "realtime" + context_length: 128000 + - model_name: "gpt-4o-mini-realtime-preview-20241217-global" + litellm_params: + model: azure/gpt-4o-mini-realtime-preview-20241217-global + base_model: "gpt-4o-realtime-preview" + api_version: 2025-04-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "sweden-central", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + mode: "realtime" + context_length: 128000 + + # ================ + # Video API Models + # ================ + - model_name: "sora-2-20251006-global" + litellm_params: + model: azure/sora-2-20251006-global + base_model: "azure/sora-2" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "sweden-central", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + - model_name: "sora-2-20251006-global" + litellm_params: + model: azure/sora-2-20251006-global + base_model: "azure/sora-2" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us-2", "azure"] + model_info: + access_groups: ["Default", "NA"] + + # ============================================ + # Australia East (ae) - Chat Completion Models + # ============================================ + - model_name: "gpt-35-turbo" + litellm_params: + model: azure/gpt-35-turbo + base_model: "gpt-3.5-turbo" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["APAC", "australia-east", "azure"] + model_info: + base_model: "gpt-3.5-turbo" + access_groups: ["Default", "APAC"] + context_length: 128000 + - model_name: "gpt-4.1-20250414-global" + litellm_params: + model: azure/gpt-4.1-20250414-global + base_model: "gpt-4" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["APAC", "australia-east", "azure"] + model_info: + base_model: "gpt-4" + access_groups: ["Default", "APAC"] + context_length: 128000 + - model_name: "gpt-4.1-mini-20250414-global" + litellm_params: + model: azure/gpt-4.1-mini-20250414-global + base_model: "gpt-4o-mini" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["APAC", "australia-east", "azure"] + model_info: + base_model: "gpt-4o-mini" + access_groups: ["Default", "APAC"] + context_length: 128000 + - model_name: "gpt-4.1-nano-20250414-global" + litellm_params: + model: azure/gpt-4.1-nano-20250414-global + base_model: "gpt-4o-mini" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["APAC", "australia-east", "azure"] + model_info: + access_groups: ["Default", "APAC"] + context_length: 128000 + - model_name: "gpt-4o-20240513-global" + litellm_params: + model: azure/gpt-4o-20240513-global + base_model: "gpt-4o" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["APAC", "australia-east", "azure"] + model_info: + access_groups: ["Default", "APAC"] + context_length: 128000 + - model_name: "gpt-4o-20241120-global" + litellm_params: + model: azure/gpt-4o-20241120-global + base_model: "gpt-4o" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["APAC", "australia-east", "azure"] + model_info: + access_groups: ["Default", "APAC"] + context_length: 128000 + - model_name: "gpt-4o-20241120-standard" + litellm_params: + model: azure/gpt-4o-20241120-standard + base_model: "gpt-4o" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["APAC", "australia-east", "azure"] + model_info: + access_groups: ["Default", "APAC"] + context_length: 128000 + - model_name: "gpt-4o-mini-20240718" + litellm_params: + model: azure/gpt-4o-mini-20240718 + base_model: "gpt-4o-mini" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["APAC", "australia-east", "azure"] + model_info: + access_groups: ["Default", "APAC"] + context_length: 128000 + - model_name: "gpt-o4-mini-20250416-global" + litellm_params: + model: azure/gpt-o4-mini-20250416-global + base_model: "o4-mini" + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["APAC", "australia-east", "azure"] + model_info: + access_groups: ["Default", "APAC"] + context_length: 200000 + - model_name: "o3-mini-20250131-global" + litellm_params: + model: azure/o3-mini-20250131-global + base_model: "o3-mini" + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["APAC", "australia-east", "azure"] + model_info: + access_groups: ["Default", "APAC"] + context_length: 200000 + + # ============================================= + # Australia East (ae) - Image Generation Models + # ============================================= + - model_name: "dall-e-3" + litellm_params: + model: azure/dall-e-3 + base_model: "dall-e-3" + api_version: 2023-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["APAC", "australia-east", "azure"] + model_info: + access_groups: ["Default", "APAC"] + mode: "image" + + # ======================================= + # Australia East (ae) - Embeddings Models + # ======================================= + - model_name: "text-embedding-3-large" + litellm_params: + model: azure/text-embedding-3-large + base_model: "text-embedding-3-large" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["APAC", "australia-east", "azure"] + model_info: + base_model: "text-embedding-3-large" + access_groups: ["Default", "APAC"] + mode: "embedding" + - model_name: "text-embedding-3-small" + litellm_params: + model: azure/text-embedding-3-small + base_model: "text-embedding-3-small" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["APAC", "australia-east", "azure"] + model_info: + access_groups: ["Default", "APAC"] + mode: "embedding" + - model_name: "text-embedding-3-small-1-global" + litellm_params: + model: azure/text-embedding-3-small-1-global + base_model: "text-embedding-3-small" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["APAC", "australia-east", "azure"] + model_info: + access_groups: ["Default", "APAC"] + mode: "embedding" + - model_name: "text-embedding-ada-002" + litellm_params: + model: azure/text-embedding-ada-002 + base_model: "text-embedding-ada-002" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["APAC", "australia-east", "azure"] + model_info: + access_groups: ["Default", "APAC"] + mode: "embedding" + + # ========================================= + # Canada East (ce) - Chat Completion Models + # ========================================= + - model_name: "gpt-35-turbo" + litellm_params: + model: azure/gpt-35-turbo + base_model: "gpt-3.5-turbo" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "canada-east", "azure"] + model_info: + base_model: "gpt-3.5-turbo" + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-4.1-20250414-global" + litellm_params: + model: azure/gpt-4.1-20250414-global + base_model: "gpt-4" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "canada-east", "azure"] + model_info: + base_model: "gpt-4" + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-4.1-mini-20250414-global" + litellm_params: + model: azure/gpt-4.1-mini-20250414-global + base_model: "gpt-4o-mini" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "canada-east", "azure"] + model_info: + base_model: "gpt-4o-mini" + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-4.1-nano-20250414-global" + litellm_params: + model: azure/gpt-4.1-nano-20250414-global + base_model: "gpt-4o-mini" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "canada-east", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-4o-20240513-global" + litellm_params: + model: azure/gpt-4o-20240513-global + base_model: "gpt-4o" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "canada-east", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-4o-20241120-global" + litellm_params: + model: azure/gpt-4o-20241120-global + base_model: "gpt-4o" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "canada-east", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-4o-mini-20240718" + litellm_params: + model: azure/gpt-4o-mini-20240718 + base_model: "gpt-4o-mini" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "canada-east", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-o4-mini-20250416-global" + litellm_params: + model: azure/gpt-o4-mini-20250416-global + base_model: "o4-mini" + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "canada-east", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 200000 + - model_name: "o1-20241207-global" + litellm_params: + model: azure/o1-20241207-global + base_model: "o1-2024-12-17" + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "canada-east", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 200000 + - model_name: "o3-mini-20250131-global" + litellm_params: + model: azure/o3-mini-20250131-global + base_model: "o3-mini" + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "canada-east", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 200000 + + # ==================================== + # Canada East (ce) - Embeddings Models + # ==================================== + - model_name: "text-embedding-3-large" + litellm_params: + model: azure/text-embedding-3-large + base_model: "text-embedding-3-large" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "canada-east", "azure"] + model_info: + base_model: "text-embedding-3-large" + access_groups: ["Default", "NA"] + mode: "embedding" + - model_name: "text-embedding-3-small" + litellm_params: + model: azure/text-embedding-3-small + base_model: "text-embedding-3-small" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "canada-east", "azure"] + model_info: + access_groups: ["Default", "NA"] + mode: "embedding" + - model_name: "text-embedding-3-small-1-global" + litellm_params: + model: azure/text-embedding-3-small-1-global + base_model: "text-embedding-3-small" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "canada-east", "azure"] + model_info: + access_groups: ["Default", "NA"] + mode: "embedding" + - model_name: "text-embedding-ada-002" + litellm_params: + model: azure/text-embedding-ada-002 + base_model: "text-embedding-ada-002" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "canada-east", "azure"] + model_info: + access_groups: ["Default", "NA"] + mode: "embedding" + + # ====================================== + # East US (eus) - Chat Completion Models + # ====================================== + - model_name: "gpt-35-turbo" + litellm_params: + model: azure/gpt-35-turbo + base_model: "gpt-3.5-turbo" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us", "azure"] + model_info: + base_model: "gpt-3.5-turbo" + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-4o-20241120-datazone" + litellm_params: + model: azure/gpt-4o-20241120-datazone + base_model: "gpt-4o" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 65536 + - model_name: "gpt-35-turbo-instruct" + litellm_params: + model: azure/gpt-35-turbo-instruct + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us", "azure"] + model_info: + base_model: "gpt-3.5-turbo-instruct" + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-4.1-20250414-datazone" + litellm_params: + model: azure/gpt-4.1-20250414-datazone + base_model: "gpt-4" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-4.1-20250414-global" + litellm_params: + model: azure/gpt-4.1-20250414-global + base_model: "gpt-4" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us", "azure"] + model_info: + base_model: "gpt-4" + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-4.1-mini-20250414-datazone" + litellm_params: + model: azure/gpt-4.1-mini-20250414-datazone + base_model: "gpt-4o-mini" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-4.1-mini-20250414-global" + litellm_params: + model: azure/gpt-4.1-mini-20250414-global + base_model: "gpt-4o-mini" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us", "azure"] + model_info: + base_model: "gpt-4o-mini" + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-4.1-nano-20250414-datazone" + litellm_params: + model: azure/gpt-4.1-nano-20250414-datazone + base_model: "gpt-4o-mini" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-4o-20240513-global" + litellm_params: + model: azure/gpt-4o-20240513-global + base_model: "gpt-4o" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "GPT-4o-20240513-standard" + litellm_params: + model: azure/GPT-4o-20240513-standard + base_model: "gpt-4o" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-4o-20240806-datazone" + litellm_params: + model: azure/gpt-4o-20240806-datazone + base_model: "gpt-4o" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-4o-20240806-standard" + litellm_params: + model: azure/gpt-4o-20240806-standard + base_model: "gpt-4o" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-4o-20241120-global" + litellm_params: + model: azure/gpt-4o-20241120-global + base_model: "gpt-4o" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-4o-20241120-standard" + litellm_params: + model: azure/gpt-4o-20241120-standard + base_model: "gpt-4o" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-4o-mini-20240718" + litellm_params: + model: azure/gpt-4o-mini-20240718 + base_model: "gpt-4o-mini" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-4o-mini-20240718-datazone" + litellm_params: + model: azure/gpt-4o-mini-20240718-datazone + base_model: "gpt-4o-mini" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-4o-mini-20240718-standard" + litellm_params: + model: azure/gpt-4o-mini-20240718-standard + base_model: "gpt-4o-mini" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-o4-mini-20250416-global" + litellm_params: + model: azure/gpt-o4-mini-20250416-global + base_model: "o4-mini" + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 200000 + - model_name: "o1-20241207-global" + litellm_params: + model: azure/o1-20241207-global + base_model: "o1-2024-12-17" + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 200000 + - model_name: "o1-mini-20240912-global" + litellm_params: + model: azure/o1-mini-20240912-global + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 200000 + - model_name: "o1-mini-20240912-standard" + litellm_params: + model: azure/o1-mini-20240912-standard + base_model: "o1-mini" + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 200000 + - model_name: "o3-20250416-datazone" + litellm_params: + model: azure/o3-20250416-datazone + base_model: "o3" + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 200000 + - model_name: "o3-mini-20250131-datazone" + litellm_params: + model: azure/o3-mini-20250131-datazone + base_model: "o3-mini" + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 200000 + - model_name: "o3-mini-20250131-global" + litellm_params: + model: azure/o3-mini-20250131-global + base_model: "o3-mini" + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 200000 + - model_name: "o4-mini-20250416-datazone" + litellm_params: + model: azure/o4-mini-20250416-datazone + base_model: "o4-mini" + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 200000 + + # ======================================= + # East US (eus) - Image Generation Models + # ======================================= + - model_name: "dall-e-3" + litellm_params: + model: azure/dall-e-3 + base_model: "dall-e-3" + api_version: 2023-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us", "azure"] + model_info: + access_groups: ["Default", "NA"] + mode: "image" + + # ================================= + # East US (eus) - Embeddings Models + # ================================= + - model_name: "text-embedding-3-large" + litellm_params: + model: azure/text-embedding-3-large + base_model: "text-embedding-3-large" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us", "azure"] + model_info: + base_model: "text-embedding-3-large" + access_groups: ["Default", "NA"] + mode: "embedding" + - model_name: "text-embedding-3-small" + litellm_params: + model: azure/text-embedding-3-small + base_model: "text-embedding-3-small" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us", "azure"] + model_info: + access_groups: ["Default", "NA"] + mode: "embedding" + - model_name: "text-embedding-3-small-1-global" + litellm_params: + model: azure/text-embedding-3-small-1-global + base_model: "text-embedding-3-small" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us", "azure"] + model_info: + access_groups: ["Default", "NA"] + mode: "embedding" + - model_name: "text-embedding-ada-002" + litellm_params: + model: azure/text-embedding-ada-002 + base_model: "text-embedding-ada-002" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "east-us", "azure"] + model_info: + access_groups: ["Default", "NA"] + mode: "embedding" + + # ============================================ + # France Central (fc) - Chat Completion Models + # ============================================ + - model_name: "gpt-35-turbo" + litellm_params: + model: azure/gpt-35-turbo + base_model: "gpt-3.5-turbo" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "france-central", "azure"] + model_info: + base_model: "gpt-3.5-turbo" + access_groups: ["Default", "EMEA"] + context_length: 128000 + - model_name: "gpt-4o-20241120-datazone" + litellm_params: + model: azure/gpt-4o-20241120-datazone + base_model: "gpt-4o" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "france-central", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 65536 + - model_name: "gpt-4.1-20250414-datazone" + litellm_params: + model: azure/gpt-4.1-20250414-datazone + base_model: "gpt-4" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "france-central", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 128000 + - model_name: "gpt-4.1-20250414-global" + litellm_params: + model: azure/gpt-4.1-20250414-global + base_model: "gpt-4" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "france-central", "azure"] + model_info: + base_model: "gpt-4" + access_groups: ["Default", "EMEA"] + context_length: 128000 + - model_name: "gpt-4.1-nano-20250414-datazone" + litellm_params: + model: azure/gpt-4.1-nano-20250414-datazone + base_model: "gpt-4o-mini" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "france-central", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 128000 + - model_name: "gpt-4.1-nano-20250414-global" + litellm_params: + model: azure/gpt-4.1-nano-20250414-global + base_model: "gpt-4o-mini" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "france-central", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 128000 + - model_name: "gpt-4o-20240513-global" + litellm_params: + model: azure/gpt-4o-20240513-global + base_model: "gpt-4o" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "france-central", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 128000 + - model_name: "gpt-4o-20240806-datazone" + litellm_params: + model: azure/gpt-4o-20240806-datazone + base_model: "gpt-4o" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "france-central", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 128000 + - model_name: "gpt-4o-20241120-global" + litellm_params: + model: azure/gpt-4o-20241120-global + base_model: "gpt-4o" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "france-central", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 128000 + - model_name: "gpt-4o-20241120-standard" + litellm_params: + model: azure/gpt-4o-20241120-standard + base_model: "gpt-4o" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "france-central", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 128000 + - model_name: "gpt-4o-mini-20240718" + litellm_params: + model: azure/gpt-4o-mini-20240718 + base_model: "gpt-4o-mini" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "france-central", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 128000 + - model_name: "gpt-4o-mini-20240718-datazone" + litellm_params: + model: azure/gpt-4o-mini-20240718-datazone + base_model: "gpt-4o-mini" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "france-central", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 128000 + - model_name: "gpt-o4-mini-20250416-global" + litellm_params: + model: azure/gpt-o4-mini-20250416-global + base_model: "o4-mini" + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "france-central", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 200000 + - model_name: "o1-20241217-datazone" + litellm_params: + model: azure/o1-20241217-datazone + base_model: "o1-2024-12-17" + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "france-central", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 200000 + - model_name: "o1-20241217-global" + litellm_params: + model: azure/o1-20241217-global + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "france-central", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 200000 + - model_name: "o3-20250416-datazone" + litellm_params: + model: azure/o3-20250416-datazone + base_model: "o3" + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "france-central", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 200000 + - model_name: "o3-mini-20250131-datazone" + litellm_params: + model: azure/o3-mini-20250131-datazone + base_model: "o3-mini" + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "france-central", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 200000 + - model_name: "o3-mini-20250131-global" + litellm_params: + model: azure/o3-mini-20250131-global + base_model: "o3-mini" + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "france-central", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 200000 + - model_name: "o4-mini-20250416-datazone" + litellm_params: + model: azure/o4-mini-20250416-datazone + base_model: "o4-mini" + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "france-central", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 200000 + + # ======================================= + # France Central (fc) - Embeddings Models + # ======================================= + - model_name: "text-embedding-3-large-1-standard" + litellm_params: + model: azure/text-embedding-3-large-1-standard + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "france-central", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + mode: "embedding" + - model_name: "text-embedding-3-small-1-global" + litellm_params: + model: azure/text-embedding-3-small-1-global + base_model: "text-embedding-3-small" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "france-central", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + mode: "embedding" + - model_name: "text-embedding-ada-002" + litellm_params: + model: azure/text-embedding-ada-002 + base_model: "text-embedding-ada-002" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "france-central", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + mode: "embedding" + + # ======================================== + # Japan East (je) - Chat Completion Models + # ======================================== + - model_name: "gpt-35-turbo" + litellm_params: + model: azure/gpt-35-turbo + base_model: "gpt-3.5-turbo" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["APAC", "japan-east", "azure"] + model_info: + base_model: "gpt-3.5-turbo" + access_groups: ["Default", "APAC"] + context_length: 128000 + - model_name: "gpt-4.1-20250414-global" + litellm_params: + model: azure/gpt-4.1-20250414-global + base_model: "gpt-4" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["APAC", "japan-east", "azure"] + model_info: + base_model: "gpt-4" + access_groups: ["Default", "APAC"] + context_length: 128000 + - model_name: "gpt-4.1-nano-20250414-global" + litellm_params: + model: azure/gpt-4.1-nano-20250414-global + base_model: "gpt-4o-mini" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["APAC", "japan-east", "azure"] + model_info: + access_groups: ["Default", "APAC"] + context_length: 128000 + - model_name: "gpt-4o-20240513-global" + litellm_params: + model: azure/gpt-4o-20240513-global + base_model: "gpt-4o" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["APAC", "japan-east", "azure"] + model_info: + access_groups: ["Default", "APAC"] + context_length: 128000 + - model_name: "gpt-4o-20241120-global" + litellm_params: + model: azure/gpt-4o-20241120-global + base_model: "gpt-4o" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["APAC", "japan-east", "azure"] + model_info: + access_groups: ["Default", "APAC"] + context_length: 128000 + - model_name: "gpt-4o-20241120-standard" + litellm_params: + model: azure/gpt-4o-20241120-standard + base_model: "gpt-4o" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["APAC", "japan-east", "azure"] + model_info: + access_groups: ["Default", "APAC"] + context_length: 128000 + - model_name: "gpt-4o-mini-20240718" + litellm_params: + model: azure/gpt-4o-mini-20240718 + base_model: "gpt-4o-mini" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["APAC", "japan-east", "azure"] + model_info: + access_groups: ["Default", "APAC"] + context_length: 128000 + - model_name: "gpt-o4-mini-20250416-global" + litellm_params: + model: azure/gpt-o4-mini-20250416-global + base_model: "o4-mini" + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["APAC", "japan-east", "azure"] + model_info: + access_groups: ["Default", "APAC"] + context_length: 200000 + - model_name: "o1-20241217-global" + litellm_params: + model: azure/o1-20241217-global + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["APAC", "japan-east", "azure"] + model_info: + access_groups: ["Default", "APAC"] + context_length: 200000 + - model_name: "o3-mini-20250131-global" + litellm_params: + model: azure/o3-mini-20250131-global + base_model: "o3-mini" + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["APAC", "japan-east", "azure"] + model_info: + access_groups: ["Default", "APAC"] + context_length: 200000 + + # =================================== + # Japan East (je) - Embeddings Models + # =================================== + - model_name: "text-embedding-3-large-1-standard" + litellm_params: + model: azure/text-embedding-3-large-1-standard + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["APAC", "japan-east", "azure"] + model_info: + access_groups: ["Default", "APAC"] + mode: "embedding" + - model_name: "text-embedding-3-small-1-global" + litellm_params: + model: azure/text-embedding-3-small-1-global + base_model: "text-embedding-3-small" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["APAC", "japan-east", "azure"] + model_info: + access_groups: ["Default", "APAC"] + mode: "embedding" + - model_name: "text-embedding-3-small-1-standard" + litellm_params: + model: azure/text-embedding-3-small-1-standard + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["APAC", "japan-east", "azure"] + model_info: + access_groups: ["Default", "APAC"] + mode: "embedding" + - model_name: "text-embedding-ada-002" + litellm_params: + model: azure/text-embedding-ada-002 + base_model: "text-embedding-ada-002" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["APAC", "japan-east", "azure"] + model_info: + access_groups: ["Default", "APAC"] + mode: "embedding" + + # ======================================= + # UK South (uks) - Chat Completion Models + # ======================================= + - model_name: "gpt-35-turbo" + litellm_params: + model: azure/gpt-35-turbo + base_model: "gpt-3.5-turbo" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "uk-south", "azure"] + model_info: + base_model: "gpt-3.5-turbo" + access_groups: ["Default", "EMEA"] + context_length: 128000 + - model_name: "gpt-4.1-20250414-global" + litellm_params: + model: azure/gpt-4.1-20250414-global + base_model: "gpt-4" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "uk-south", "azure"] + model_info: + base_model: "gpt-4" + access_groups: ["Default", "EMEA"] + context_length: 128000 + - model_name: "gpt-4.1-mini-20250414-global" + litellm_params: + model: azure/gpt-4.1-mini-20250414-global + base_model: "gpt-4o-mini" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "uk-south", "azure"] + model_info: + base_model: "gpt-4o-mini" + access_groups: ["Default", "EMEA"] + context_length: 128000 + - model_name: "gpt-4.1-nano-20250414-global" + litellm_params: + model: azure/gpt-4.1-nano-20250414-global + base_model: "gpt-4o-mini" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "uk-south", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 128000 + - model_name: "gpt-4o-20240513-global" + litellm_params: + model: azure/gpt-4o-20240513-global + base_model: "gpt-4o" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "uk-south", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 128000 + - model_name: "gpt-4o-20241120-global" + litellm_params: + model: azure/gpt-4o-20241120-global + base_model: "gpt-4o" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "uk-south", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 128000 + - model_name: "gpt-4o-20241120-standard" + litellm_params: + model: azure/gpt-4o-20241120-standard + base_model: "gpt-4o" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "uk-south", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 128000 + - model_name: "gpt-4o-mini-20240718" + litellm_params: + model: azure/gpt-4o-mini-20240718 + base_model: "gpt-4o-mini" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "uk-south", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 128000 + - model_name: "gpt-o4-mini-20250416-global" + litellm_params: + model: azure/gpt-o4-mini-20250416-global + base_model: "o4-mini" + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "uk-south", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 200000 + - model_name: "o1-20241217-global" + litellm_params: + model: azure/o1-20241217-global + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "uk-south", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 200000 + - model_name: "o3-mini-20250131-global" + litellm_params: + model: azure/o3-mini-20250131-global + base_model: "o3-mini" + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "uk-south", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 200000 + + # ================================== + # UK South (uks) - Embeddings Models + # ================================== + - model_name: "text-embedding-3-large-1-standard" + litellm_params: + model: azure/text-embedding-3-large-1-standard + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "uk-south", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + mode: "embedding" + - model_name: "text-embedding-3-small-1-global" + litellm_params: + model: azure/text-embedding-3-small-1-global + base_model: "text-embedding-3-small" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "uk-south", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + mode: "embedding" + - model_name: "text-embedding-ada-002" + litellm_params: + model: azure/text-embedding-ada-002 + base_model: "text-embedding-ada-002" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "uk-south", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + mode: "embedding" + + # ========================================= + # West Europe (we) - Chat Completion Models + # ========================================= + - model_name: "gpt-35-turbo" + litellm_params: + model: azure/gpt-35-turbo + base_model: "gpt-3.5-turbo" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "west-europe", "azure"] + model_info: + base_model: "gpt-3.5-turbo" + access_groups: ["Default", "EMEA"] + context_length: 128000 + - model_name: "gpt-4o-20241120-datazone" + litellm_params: + model: azure/gpt-4o-20241120-datazone + base_model: "gpt-4o" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "west-europe", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 65536 + - model_name: "gpt-4.1-20250414-datazone" + litellm_params: + model: azure/gpt-4.1-20250414-datazone + base_model: "gpt-4" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "west-europe", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 128000 + - model_name: "gpt-4.1-20250414-global" + litellm_params: + model: azure/gpt-4.1-20250414-global + base_model: "gpt-4" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "west-europe", "azure"] + model_info: + base_model: "gpt-4" + access_groups: ["Default", "EMEA"] + context_length: 128000 + - model_name: "gpt-4.1-mini-20250414-global" + litellm_params: + model: azure/gpt-4.1-mini-20250414-global + base_model: "gpt-4o-mini" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "west-europe", "azure"] + model_info: + base_model: "gpt-4o-mini" + access_groups: ["Default", "EMEA"] + context_length: 128000 + - model_name: "gpt-4.1-nano-20250414-datazone" + litellm_params: + model: azure/gpt-4.1-nano-20250414-datazone + base_model: "gpt-4o-mini" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "west-europe", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 128000 + - model_name: "gpt-4.1-nano-20250414-global" + litellm_params: + model: azure/gpt-4.1-nano-20250414-global + base_model: "gpt-4o-mini" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "west-europe", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 128000 + - model_name: "gpt-4o-20240513-global" + litellm_params: + model: azure/gpt-4o-20240513-global + base_model: "gpt-4o" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "west-europe", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 128000 + - model_name: "gpt-4o-20240806-datazone" + litellm_params: + model: azure/gpt-4o-20240806-datazone + base_model: "gpt-4o" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "west-europe", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 128000 + - model_name: "gpt-4o-20241120-global" + litellm_params: + model: azure/gpt-4o-20241120-global + base_model: "gpt-4o" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "west-europe", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 128000 + - model_name: "gpt-4o-mini-20240718" + litellm_params: + model: azure/gpt-4o-mini-20240718 + base_model: "gpt-4o-mini" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "west-europe", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 128000 + - model_name: "gpt-o4-mini-20250416-global" + litellm_params: + model: azure/gpt-o4-mini-20250416-global + base_model: "o4-mini" + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "west-europe", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 200000 + - model_name: "o1-20241217-datazone" + litellm_params: + model: azure/o1-20241217-datazone + base_model: "o1-2024-12-17" + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "west-europe", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 200000 + - model_name: "o1-20241217-global" + litellm_params: + model: azure/o1-20241217-global + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "west-europe", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 200000 + - model_name: "o3-20250416-datazone" + litellm_params: + model: azure/o3-20250416-datazone + base_model: "o3" + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "west-europe", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 200000 + - model_name: "o3-mini-20250131-datazone" + litellm_params: + model: azure/o3-mini-20250131-datazone + base_model: "o3-mini" + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "west-europe", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 200000 + - model_name: "o3-mini-20250131-global" + litellm_params: + model: azure/o3-mini-20250131-global + base_model: "o3-mini" + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "west-europe", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 200000 + - model_name: "o4-mini-20250416-datazone" + litellm_params: + model: azure/o4-mini-20250416-datazone + base_model: "o4-mini" + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "west-europe", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 200000 + + # =============================== + # West Europe (we) - Audio Models + # =============================== + - model_name: "whisper-1" + litellm_params: + model: azure/whisper + base_model: "whisper-1" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "west-europe", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + mode: "audio" + + # ==================================== + # West Europe (we) - Embeddings Models + # ==================================== + - model_name: "text-embedding-3-small-1-global" + litellm_params: + model: azure/text-embedding-3-small-1-global + base_model: "text-embedding-3-small" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "west-europe", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + mode: "embedding" + - model_name: "text-embedding-ada-002" + litellm_params: + model: azure/text-embedding-ada-002 + base_model: "text-embedding-ada-002" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "west-europe", "azure"] + model_info: + access_groups: ["Default", "EMEA"] + mode: "embedding" + + # ====================================== + # West US (wus) - Chat Completion Models + # ====================================== + - model_name: "gpt-35-turbo" + litellm_params: + model: azure/gpt-35-turbo + base_model: "gpt-3.5-turbo" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "west-us", "azure"] + model_info: + base_model: "gpt-3.5-turbo" + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-4o-20241120-datazone" + litellm_params: + model: azure/gpt-4o-20241120-datazone + base_model: "gpt-4o" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "west-us", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 65536 + - model_name: "gpt-4.1-20250414-datazone" + litellm_params: + model: azure/gpt-4.1-20250414-datazone + base_model: "gpt-4" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "west-us", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-4.1-20250414-global" + litellm_params: + model: azure/gpt-4.1-20250414-global + base_model: "gpt-4" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "west-us", "azure"] + model_info: + base_model: "gpt-4" + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-4.1-mini-20250414-datazone" + litellm_params: + model: azure/gpt-4.1-mini-20250414-datazone + base_model: "gpt-4o-mini" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "west-us", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-4.1-mini-20250414-global" + litellm_params: + model: azure/gpt-4.1-mini-20250414-global + base_model: "gpt-4o-mini" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "west-us", "azure"] + model_info: + base_model: "gpt-4o-mini" + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-4.1-nano-20250414-datazone" + litellm_params: + model: azure/gpt-4.1-nano-20250414-datazone + base_model: "gpt-4o-mini" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "west-us", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-4.1-nano-20250414-global" + litellm_params: + model: azure/gpt-4.1-nano-20250414-global + base_model: "gpt-4o-mini" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "west-us", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-4o-2024-08-06-Standard" + litellm_params: + model: azure/gpt-4o-2024-08-06-Standard + base_model: "gpt-4o" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "west-us", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-4o-20240513-global" + litellm_params: + model: azure/gpt-4o-20240513-global + base_model: "gpt-4o" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "west-us", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "GPT-4o-20240513-standard" + litellm_params: + model: azure/GPT-4o-20240513-standard + base_model: "gpt-4o" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "west-us", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-4o-20240806-datazone" + litellm_params: + model: azure/gpt-4o-20240806-datazone + base_model: "gpt-4o" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "west-us", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-4o-20241120-global" + litellm_params: + model: azure/gpt-4o-20241120-global + base_model: "gpt-4o" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "west-us", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-4o-20241120-standard" + litellm_params: + model: azure/gpt-4o-20241120-standard + base_model: "gpt-4o" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "west-us", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-4o-mini-20240718" + litellm_params: + model: azure/gpt-4o-mini-20240718 + base_model: "gpt-4o-mini" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "west-us", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-4o-mini-20240718-datazone" + litellm_params: + model: azure/gpt-4o-mini-20240718-datazone + base_model: "gpt-4o-mini" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "west-us", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-4o-mini-20240718-standard" + litellm_params: + model: azure/gpt-4o-mini-20240718-standard + base_model: "gpt-4o-mini" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "west-us", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 128000 + - model_name: "gpt-o4-mini-20250416-global" + litellm_params: + model: azure/gpt-o4-mini-20250416-global + base_model: "o4-mini" + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "west-us", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 200000 + - model_name: "o1-20241207-global" + litellm_params: + model: azure/o1-20241207-global + base_model: "o1-2024-12-17" + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "west-us", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 200000 + - model_name: "o1-mini-20240912-global" + litellm_params: + model: azure/o1-mini-20240912-global + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "west-us", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 200000 + - model_name: "o1-mini-20240912-standard" + litellm_params: + model: azure/o1-mini-20240912-standard + base_model: "o1-mini" + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "west-us", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 200000 + - model_name: "o3-20250416-datazone" + litellm_params: + model: azure/o3-20250416-datazone + base_model: "o3" + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "west-us", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 200000 + - model_name: "o3-mini-20250131-datazone" + litellm_params: + model: azure/o3-mini-20250131-datazone + base_model: "o3-mini" + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "west-us", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 200000 + - model_name: "o3-mini-20250131-global" + litellm_params: + model: azure/o3-mini-20250131-global + base_model: "o3-mini" + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "west-us", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 200000 + - model_name: "o4-mini-20250416-datazone" + litellm_params: + model: azure/o4-mini-20250416-datazone + base_model: "o4-mini" + api_version: 2024-12-01-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "west-us", "azure"] + model_info: + access_groups: ["Default", "NA"] + context_length: 200000 + + # ================================= + # West US (wus) - Embeddings Models + # ================================= + - model_name: "text-embedding-3-small-1-global" + litellm_params: + model: azure/text-embedding-3-small-1-global + base_model: "text-embedding-3-small" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "west-us", "azure"] + model_info: + access_groups: ["Default", "NA"] + mode: "embedding" + - model_name: "text-embedding-3-small-1-standard" + litellm_params: + model: azure/text-embedding-3-small-1-standard + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "west-us", "azure"] + model_info: + access_groups: ["Default", "NA"] + mode: "embedding" + - model_name: "text-embedding-ada-002" + litellm_params: + model: azure/text-embedding-ada-002 + base_model: "text-embedding-ada-002" + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "west-us", "azure"] + model_info: + access_groups: ["Default", "NA"] + mode: "embedding" + + # ===================== + # West US3 (wus3) - PTU + # ===================== + - model_name: "GPT-4o-2024-05-13-PTU" + litellm_params: + model: azure/GPT-4o-2024-05-13-PTU + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["NA", "west-us-3", "azure"] + model_info: + access_groups: ["PTU"] + context_length: 128000 + + # =========================== + # AWS Frankfurt (fr) - Models + # =========================== + - model_name: eu.anthropic.claude-haiku-4-5-20251001-v1:0 + litellm_params: + model: bedrock/eu.anthropic.claude-haiku-4-5-20251001-v1:0 + aws_access_key_id: dummy-aws-access-key + aws_secret_access_key: dummy-aws-secret-key + api_base: *api_base + tags: ["EMEA", "frankfurt", "aws"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 200000 + - model_name: amazon.titan-text-express-v1 + litellm_params: + model: bedrock/amazon.titan-text-express-v1 + aws_access_key_id: dummy-aws-access-key + aws_secret_access_key: dummy-aws-secret-key + api_base: *api_base + tags: ["EMEA", "frankfurt", "aws"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 200000 + - model_name: eu.anthropic.claude-sonnet-4-5-20250929-v1:0 + litellm_params: + model: bedrock/eu.anthropic.claude-sonnet-4-5-20250929-v1:0 + aws_access_key_id: dummy-aws-access-key + aws_secret_access_key: dummy-aws-secret-key + api_base: *api_base + tags: ["EMEA", "frankfurt", "aws"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 200000 + - model_name: global.anthropic.claude-sonnet-4-5-20250929-v1:0 + litellm_params: + model: bedrock/global.anthropic.claude-sonnet-4-5-20250929-v1:0 + aws_access_key_id: dummy-aws-access-key + aws_secret_access_key: dummy-aws-secret-key + api_base: *api_base + tags: ["EMEA", "frankfurt", "aws"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 200000 + - model_name: amazon.titan-embed-text-v1 + litellm_params: + model: bedrock/amazon.titan-embed-text-v1 + aws_access_key_id: dummy-aws-access-key + aws_secret_access_key: dummy-aws-secret-key + api_base: *api_base + tags: ["EMEA", "frankfurt", "aws"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 200000 + - model_name: anthropic.claude-3-sonnet-20240229-v1:0 + litellm_params: + model: bedrock/anthropic.claude-3-sonnet-20240229-v1:0 + aws_access_key_id: dummy-aws-access-key + aws_secret_access_key: dummy-aws-secret-key + api_base: *api_base + tags: ["EMEA", "frankfurt", "aws"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 200000 + - model_name: anthropic.claude-3-haiku-20240307-v1:0 + litellm_params: + model: bedrock/anthropic.claude-3-haiku-20240307-v1:0 + aws_access_key_id: dummy-aws-access-key + aws_secret_access_key: dummy-aws-secret-key + api_base: *api_base + tags: ["EMEA", "frankfurt", "aws"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 200000 + - model_name: anthropic.claude-3-5-sonnet-20240620-v1:0 + litellm_params: + model: bedrock/anthropic.claude-3-5-sonnet-20240620-v1:0 + aws_access_key_id: dummy-aws-access-key + aws_secret_access_key: dummy-aws-secret-key + api_base: *api_base + tags: ["EMEA", "frankfurt", "aws"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 200000 + - model_name: eu.anthropic.claude-3-7-sonnet-20250219-v1:0 + litellm_params: + model: bedrock/eu.anthropic.claude-3-7-sonnet-20250219-v1:0 + aws_access_key_id: dummy-aws-access-key + aws_secret_access_key: dummy-aws-secret-key + api_base: *api_base + tags: ["EMEA", "frankfurt", "aws"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 200000 + - model_name: eu.anthropic.claude-sonnet-4-20250514-v1:0 + litellm_params: + model: bedrock/eu.anthropic.claude-sonnet-4-20250514-v1:0 + aws_access_key_id: dummy-aws-access-key + aws_secret_access_key: dummy-aws-secret-key + api_base: *api_base + tags: ["EMEA", "frankfurt", "aws"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 200000 + + # ======================================== + # Ireland (eu-west-1) - AWS Bedrock Models + # ======================================== + - model_name: anthropic.claude-3-sonnet-20240229-v1:0 + litellm_params: + model: bedrock/anthropic.claude-3-sonnet-20240229-v1:0 + aws_access_key_id: dummy-aws-access-key + aws_secret_access_key: dummy-aws-secret-key + api_base: *api_base + tags: ["EMEA", "ireland", "aws"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 200000 + - model_name: eu.anthropic.claude-sonnet-4-5-20250929-v1:0 + litellm_params: + model: bedrock/eu.anthropic.claude-sonnet-4-5-20250929-v1:0 + aws_access_key_id: dummy-aws-access-key + aws_secret_access_key: dummy-aws-secret-key + api_base: *api_base + tags: ["EMEA", "ireland", "aws"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 200000 + - model_name: global.anthropic.claude-sonnet-4-5-20250929-v1:0 + litellm_params: + model: bedrock/global.anthropic.claude-sonnet-4-5-20250929-v1:0 + aws_access_key_id: dummy-aws-access-key + aws_secret_access_key: dummy-aws-secret-key + api_base: *api_base + tags: ["EMEA", "ireland", "aws"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 200000 + - model_name: anthropic.claude-3-haiku-20240307-v1:0 + litellm_params: + model: bedrock/anthropic.claude-3-haiku-20240307-v1:0 + aws_access_key_id: dummy-aws-access-key + aws_secret_access_key: dummy-aws-secret-key + api_base: *api_base + tags: ["EMEA", "ireland", "aws"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 200000 + - model_name: eu.anthropic.claude-3-7-sonnet-20250219-v1:0 + litellm_params: + model: bedrock/eu.anthropic.claude-3-7-sonnet-20250219-v1:0 + aws_access_key_id: dummy-aws-access-key + aws_secret_access_key: dummy-aws-secret-key + api_base: *api_base + tags: ["EMEA", "ireland", "aws"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 200000 + - model_name: mistral.mistral-large-2402-v1:0 + litellm_params: + model: bedrock/mistral.mistral-large-2402-v1:0 + aws_access_key_id: dummy-aws-access-key + aws_secret_access_key: dummy-aws-secret-key + api_base: *api_base + tags: ["EMEA", "ireland", "aws"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 200000 + - model_name: mistral.mistral-7b-instruct-v0:2 + litellm_params: + model: bedrock/mistral.mistral-7b-instruct-v0:2 + aws_access_key_id: dummy-aws-access-key + aws_secret_access_key: dummy-aws-secret-key + api_base: *api_base + tags: ["EMEA", "ireland", "aws"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 200000 + - model_name: mistral.mixtral-8x7b-instruct-v0:1 + litellm_params: + model: bedrock/mistral.mixtral-8x7b-instruct-v0:1 + aws_access_key_id: dummy-aws-access-key + aws_secret_access_key: dummy-aws-secret-key + api_base: *api_base + tags: ["EMEA", "ireland", "aws"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 200000 + - model_name: eu.anthropic.claude-sonnet-4-20250514-v1:0 + litellm_params: + model: bedrock/eu.anthropic.claude-sonnet-4-20250514-v1:0 + aws_access_key_id: dummy-aws-access-key + aws_secret_access_key: dummy-aws-secret-key + api_base: *api_base + tags: ["EMEA", "ireland", "aws"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 200000 + + # ============================================ + # Sydney (ap-southeast-2) - AWS Bedrock Models + # ============================================ + - model_name: au.anthropic.claude-haiku-4-5-20251001-v1:0 + litellm_params: + model: bedrock/au.anthropic.claude-haiku-4-5-20251001-v1:0 + aws_access_key_id: dummy-aws-access-key + aws_secret_access_key: dummy-aws-secret-key + api_base: *api_base + tags: ["APAC", "sydney", "aws"] + model_info: + access_groups: ["Default", "APAC"] + context_length: 200000 + - model_name: anthropic.claude-3-haiku-20240307-v1:0 + litellm_params: + model: bedrock/anthropic.claude-3-haiku-20240307-v1:0 + aws_access_key_id: dummy-aws-access-key + aws_secret_access_key: dummy-aws-secret-key + api_base: *api_base + tags: ["APAC", "sydney", "aws"] + model_info: + access_groups: ["Default", "APAC"] + context_length: 200000 + - model_name: anthropic.claude-3-sonnet-20240229-v1:0 + litellm_params: + model: bedrock/anthropic.claude-3-sonnet-20240229-v1:0 + aws_access_key_id: dummy-aws-access-key + aws_secret_access_key: dummy-aws-secret-key + api_base: *api_base + tags: ["APAC", "sydney", "aws"] + model_info: + access_groups: ["Default", "APAC"] + context_length: 200000 + - model_name: amazon.titan-embed-text-v1 + litellm_params: + model: bedrock/amazon.titan-embed-text-v1 + aws_access_key_id: dummy-aws-access-key + aws_secret_access_key: dummy-aws-secret-key + api_base: *api_base + tags: ["APAC", "sydney", "aws"] + model_info: + access_groups: ["Default", "APAC"] + context_length: 200000 + - model_name: amazon.titan-text-express-v1 + litellm_params: + model: bedrock/amazon.titan-text-express-v1 + aws_access_key_id: dummy-aws-access-key + aws_secret_access_key: dummy-aws-secret-key + api_base: *api_base + tags: ["APAC", "sydney", "aws"] + model_info: + access_groups: ["Default", "APAC"] + context_length: 200000 + - model_name: cohere.embed-english-v3 + litellm_params: + model: bedrock/cohere.embed-english-v3 + aws_access_key_id: dummy-aws-access-key + aws_secret_access_key: dummy-aws-secret-key + api_base: *api_base + tags: ["APAC", "sydney", "aws"] + model_info: + access_groups: ["Default", "APAC"] + context_length: 200000 + - model_name: cohere.embed-multilingual-v3 + litellm_params: + model: bedrock/cohere.embed-multilingual-v3 + aws_access_key_id: dummy-aws-access-key + aws_secret_access_key: dummy-aws-secret-key + api_base: *api_base + tags: ["APAC", "sydney", "aws"] + model_info: + access_groups: ["Default", "APAC"] + context_length: 200000 + - model_name: amazon.titan-embed-text-v2:0 + litellm_params: + model: bedrock/amazon.titan-embed-text-v2:0 + aws_access_key_id: dummy-aws-access-key + aws_secret_access_key: dummy-aws-secret-key + api_base: *api_base + tags: ["APAC", "sydney", "aws"] + model_info: + access_groups: ["Default", "APAC"] + context_length: 200000 + - model_name: anthropic.claude-3-5-sonnet-20241022-v2:0 + litellm_params: + model: bedrock/anthropic.claude-3-5-sonnet-20241022-v2:0 + aws_access_key_id: dummy-aws-access-key + aws_secret_access_key: dummy-aws-secret-key + api_base: *api_base + tags: ["APAC", "sydney", "aws"] + model_info: + access_groups: ["Default", "APAC"] + context_length: 200000 + - model_name: mistral.mistral-large-2402-v1:0 + litellm_params: + model: bedrock/mistral.mistral-large-2402-v1:0 + aws_access_key_id: dummy-aws-access-key + aws_secret_access_key: dummy-aws-secret-key + api_base: *api_base + tags: ["APAC", "sydney", "aws"] + model_info: + access_groups: ["Default", "APAC"] + context_length: 200000 + - model_name: mistral.mistral-7b-instruct-v0:2 + litellm_params: + model: bedrock/mistral.mistral-7b-instruct-v0:2 + aws_access_key_id: dummy-aws-access-key + aws_secret_access_key: dummy-aws-secret-key + api_base: *api_base + tags: ["APAC", "sydney", "aws"] + model_info: + access_groups: ["Default", "APAC"] + context_length: 200000 + - model_name: mistral.mixtral-8x7b-instruct-v0:1 + litellm_params: + model: bedrock/mistral.mixtral-8x7b-instruct-v0:1 + aws_access_key_id: dummy-aws-access-key + aws_secret_access_key: dummy-aws-secret-key + api_base: *api_base + tags: ["APAC", "sydney", "aws"] + model_info: + access_groups: ["Default", "APAC"] + context_length: 200000 + + # =============================================== + # North Virginia (us-east-1) - AWS Bedrock Models + # =============================================== + - model_name: us.anthropic.claude-haiku-4-5-20251001-v1:0 + litellm_params: + model: bedrock/us.anthropic.claude-haiku-4-5-20251001-v1:0 + aws_access_key_id: dummy-aws-access-key + aws_secret_access_key: dummy-aws-secret-key + api_base: *api_base + tags: ["NA", "north-virginia", "aws"] + model_info: + access_groups: ["Default", "NA"] + context_length: 200000 + - model_name: amazon.titan-text-express-v1 + litellm_params: + model: bedrock/amazon.titan-text-express-v1 + aws_access_key_id: dummy-aws-access-key + aws_secret_access_key: dummy-aws-secret-key + api_base: *api_base + tags: ["NA", "north-virginia", "aws"] + model_info: + access_groups: ["Default", "NA"] + context_length: 200000 + - model_name: us.anthropic.claude-sonnet-4-5-20250929-v1:0 + litellm_params: + model: bedrock/us.anthropic.claude-sonnet-4-5-20250929-v1:0 + aws_access_key_id: dummy-aws-access-key + aws_secret_access_key: dummy-aws-secret-key + api_base: *api_base + tags: ["NA", "north-virginia", "aws"] + model_info: + access_groups: ["Default", "NA"] + context_length: 200000 + - model_name: global.anthropic.claude-sonnet-4-5-20250929-v1:0 + litellm_params: + model: bedrock/global.anthropic.claude-sonnet-4-5-20250929-v1:0 + aws_access_key_id: dummy-aws-access-key + aws_secret_access_key: dummy-aws-secret-key + api_base: *api_base + tags: ["NA", "north-virginia", "aws"] + model_info: + access_groups: ["Default", "NA"] + context_length: 200000 + - model_name: amazon.titan-embed-text-v1 + litellm_params: + model: bedrock/amazon.titan-embed-text-v1 + aws_access_key_id: dummy-aws-access-key + aws_secret_access_key: dummy-aws-secret-key + api_base: *api_base + tags: ["NA", "north-virginia", "aws"] + model_info: + access_groups: ["Default", "NA"] + context_length: 200000 + - model_name: anthropic.claude-3-sonnet-20240229-v1:0 + litellm_params: + model: bedrock/anthropic.claude-3-sonnet-20240229-v1:0 + aws_access_key_id: dummy-aws-access-key + aws_secret_access_key: dummy-aws-secret-key + api_base: *api_base + tags: ["NA", "north-virginia", "aws"] + model_info: + access_groups: ["Default", "NA"] + context_length: 200000 + - model_name: anthropic.claude-3-haiku-20240307-v1:0 + litellm_params: + model: bedrock/anthropic.claude-3-haiku-20240307-v1:0 + aws_access_key_id: dummy-aws-access-key + aws_secret_access_key: dummy-aws-secret-key + api_base: *api_base + tags: ["NA", "north-virginia", "aws"] + model_info: + access_groups: ["Default", "NA"] + context_length: 200000 + - model_name: anthropic.claude-3-5-sonnet-20240620-v1:0 + litellm_params: + model: bedrock/anthropic.claude-3-5-sonnet-20240620-v1:0 + aws_access_key_id: dummy-aws-access-key + aws_secret_access_key: dummy-aws-secret-key + api_base: *api_base + tags: ["NA", "north-virginia", "aws"] + model_info: + access_groups: ["Default", "NA"] + context_length: 200000 + - model_name: us.anthropic.claude-3-5-sonnet-20241022-v2:0 + litellm_params: + model: bedrock/us.anthropic.claude-3-5-sonnet-20241022-v2:0 + aws_access_key_id: dummy-aws-access-key + aws_secret_access_key: dummy-aws-secret-key + api_base: *api_base + tags: ["NA", "north-virginia", "aws"] + model_info: + access_groups: ["Default", "NA"] + context_length: 200000 + - model_name: us.anthropic.claude-3-7-sonnet-20250219-v1:0 + litellm_params: + model: bedrock/us.anthropic.claude-3-7-sonnet-20250219-v1:0 + aws_access_key_id: dummy-aws-access-key + aws_secret_access_key: dummy-aws-secret-key + api_base: *api_base + tags: ["NA", "north-virginia", "aws"] + model_info: + access_groups: ["Default", "NA"] + context_length: 200000 + - model_name: us.deepseek.r1-v1:0 + litellm_params: + model: bedrock/us.deepseek.r1-v1:0 + aws_access_key_id: dummy-aws-access-key + aws_secret_access_key: dummy-aws-secret-key + api_base: *api_base + tags: ["NA", "north-virginia", "aws"] + model_info: + access_groups: ["Default", "NA"] + context_length: 200000 + - model_name: mistral.mistral-large-2402-v1:0 + litellm_params: + model: bedrock/mistral.mistral-large-2402-v1:0 + aws_access_key_id: dummy-aws-access-key + aws_secret_access_key: dummy-aws-secret-key + api_base: *api_base + tags: ["NA", "north-virginia", "aws"] + model_info: + access_groups: ["Default", "NA"] + context_length: 200000 + - model_name: mistral.mistral-7b-instruct-v0:2 + litellm_params: + model: bedrock/mistral.mistral-7b-instruct-v0:2 + aws_access_key_id: dummy-aws-access-key + aws_secret_access_key: dummy-aws-secret-key + api_base: *api_base + tags: ["NA", "north-virginia", "aws"] + model_info: + access_groups: ["Default", "NA"] + context_length: 200000 + - model_name: mistral.mixtral-8x7b-instruct-v0:1 + litellm_params: + model: bedrock/mistral.mixtral-8x7b-instruct-v0:1 + aws_access_key_id: dummy-aws-access-key + aws_secret_access_key: dummy-aws-secret-key + api_base: *api_base + tags: ["NA", "north-virginia", "aws"] + model_info: + access_groups: ["Default", "NA"] + context_length: 200000 + - model_name: mistral.mistral-small-2402-v1:0 + litellm_params: + model: bedrock/mistral.mistral-small-2402-v1:0 + aws_access_key_id: dummy-aws-access-key + aws_secret_access_key: dummy-aws-secret-key + api_base: *api_base + tags: ["NA", "north-virginia", "aws"] + model_info: + access_groups: ["Default", "NA"] + context_length: 200000 + - model_name: us.anthropic.claude-sonnet-4-20250514-v1:0 + litellm_params: + model: bedrock/us.anthropic.claude-sonnet-4-20250514-v1:0 + aws_access_key_id: dummy-aws-access-key + aws_secret_access_key: dummy-aws-secret-key + api_base: *api_base + tags: ["NA", "north-virginia", "aws"] + model_info: + access_groups: ["Default", "NA"] + context_length: 200000 + - model_name: us.anthropic.claude-opus-4-20250514-v1:0 + litellm_params: + model: bedrock/us.anthropic.claude-opus-4-20250514-v1:0 + aws_access_key_id: dummy-aws-access-key + aws_secret_access_key: dummy-aws-secret-key + api_base: *api_base + tags: ["NA", "north-virginia", "aws"] + model_info: + access_groups: ["Default", "NA"] + context_length: 200000 + + # =========================== + # Global - AWS Bedrock Models + # =========================== + - model_name: global.anthropic.claude-haiku-4-5-20251001-v1:0 + litellm_params: + model: bedrock/global.anthropic.claude-haiku-4-5-20251001-v1:0 + aws_access_key_id: dummy-aws-access-key + aws_secret_access_key: dummy-aws-secret-key + api_base: *api_base + tags: ["Default", "frankfurt", "aws"] + model_info: + access_groups: ["Default", "EMEA", "APAC", "NA"] + context_length: 200000 + +################################# +### Models - Google Gemini (Europe West) +################################# + + # ================================== + # Europe West - Google Gemini Models + # ================================== + - model_name: embedding-001 + litellm_params: + model: gemini/embedding-001 + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "europe-west", "gcp"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 4000000 + - model_name: text-embedding-004 + litellm_params: + model: gemini/text-embedding-004 + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "europe-west", "gcp"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 4000000 + - model_name: gemini-2.0-flash-001 + litellm_params: + model: gemini/gemini-2.0-flash-001 + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "europe-west", "gcp"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 4000000 + - model_name: gemini-2.0-flash-lite + litellm_params: + model: gemini/gemini-2.0-flash-lite + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "europe-west", "gcp"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 4000000 + - model_name: gemini-2.5-pro + litellm_params: + model: gemini/gemini-2.5-pro + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "europe-west", "gcp"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 4000000 + - model_name: gemini-2.5-flash + litellm_params: + model: gemini/gemini-2.5-flash + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "europe-west", "gcp"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 4000000 + - model_name: gemini-3-pro-preview + litellm_params: + model: gemini/gemini-3-pro-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "europe-west", "gcp"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 4000000 + - model_name: gemini-3-flash-preview + litellm_params: + model: gemini/gemini-3-flash-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "europe-west", "gcp"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 4000000 + - model_name: gemini-3-pro-image-preview + litellm_params: + model: gemini/gemini-3-pro-image-preview + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "europe-west", "gcp"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 4000000 + + # ============================ + # Video - Google Gemini Models + # ============================ + - model_name: veo-3.1-generate-001 + litellm_params: + model: gemini/veo-3.1-generate-001 + api_key: REDACTED_MODEL_API_KEY + api_base: *api_base + tags: ["EMEA", "europe-west", "gcp"] + model_info: + access_groups: ["Default", "EMEA"] + context_length: 4000000 + +################################# +### Environment Variables +################################# + +environment_variables: + SLACK_WEBHOOK_URL: "REDACTED" # set via env or replace for alerting + # # Langfuse credentials (required for langfuse callback) + # LANGFUSE_PUBLIC_KEY: "REDACTED" + # LANGFUSE_SECRET_KEY: "REDACTED" + # LANGFUSE_HOST: "https://us.cloud.langfuse.com"