Merge pull request #41282 from BerriAI/litellm_fast_mode_toggle_0915

feat(auto-router): add per-model Fast mode toggle

litellm/llms/anthropic/common_utils.py

@@ -539,6 +539,13 @@ class AnthropicModelInfo(BaseLLMModelInfo):
         value: Final = litellm.model_cost.get(model, {}).get(key)
         return value if isinstance(value, bool) else None
 
+    @staticmethod
+    def supports_fast_mode(model: str, custom_llm_provider: str) -> bool:
+        return (
+            custom_llm_provider == "anthropic"
+            and AnthropicModelInfo._get_exact_model_capability(model, "supports_fast_mode") is True
+        )
+
     @staticmethod
     def _get_provider_resolved_capability(model: str, key: str, custom_llm_provider: str) -> bool | None:
         """Resolve boolean capability ``key`` for ``model`` under the caller's provider.

litellm/model_prices_and_context_window_backup.json

@@ -14062,6 +14062,7 @@
         },
         "supports_output_config": true,
         "supports_speed": true,
+        "supports_fast_mode": true,
         "prompt_cache_min_tokens": 512,
         "source": "https://platform.claude.com/docs/en/about-claude/pricing"
     },
@@ -14103,6 +14104,7 @@
         },
         "supports_output_config": true,
         "supports_speed": true,
+        "supports_fast_mode": true,
         "prompt_cache_min_tokens": 1024,
         "source": "https://platform.claude.com/docs/en/about-claude/pricing"
     },

litellm/router.py

@@ -102,6 +102,7 @@ from litellm.litellm_core_utils.sensitive_data_masker import (
     mask_sensitive_structure,
 )
 from litellm.litellm_core_utils.token_counter import offload_token_count
+from litellm.llms.anthropic.common_utils import AnthropicModelInfo
 from litellm.llms.base_llm.passthrough.transformation import replace_path_segment
 from litellm.llms.base_llm.vector_store.transformation import (
     RouterVectorStoreEmbeddingExecutor,
@@ -10818,6 +10819,7 @@ class Router:
                         "model_group": user_facing_model_group_name,
                         "providers": [llm_provider],
                         **model_info,
+                        "supports_fast_mode": True,
                         "supported_reasoning_efforts": None,
                     }
                 )
@@ -10896,6 +10898,9 @@ class Router:
                 if model_info.get("rpm", None) is not None and _deployment_rpm is None:
                     _deployment_rpm = model_info.get("rpm")
 
+            model_group_info.supports_fast_mode = model_group_info.supports_fast_mode and (
+                AnthropicModelInfo.supports_fast_mode(litellm_model, llm_provider)
+            )
             deployment_reasoning_efforts = (
                 resolve_supported_reasoning_efforts(  # rebind-ok: recalculated per deployment
                     model_info, deployment_is_mapped=deployment_is_mapped

litellm/types/router.py

@@ -722,6 +722,7 @@ class ModelGroupInfo(BaseModel):
     supports_url_context: bool = Field(default=False)
     supports_reasoning: bool = Field(default=False)
     supports_function_calling: bool = Field(default=False)
+    supports_fast_mode: bool = Field(default=False)
     supported_reasoning_efforts: tuple[str, ...] | None = Field(default=None)
     supported_openai_params: list[str] | None = Field(default=[])
     configurable_clientside_auth_params: CONFIGURABLE_CLIENTSIDE_AUTH_PARAMS = None

model_prices_and_context_window.json

@@ -14062,6 +14062,7 @@
         },
         "supports_output_config": true,
         "supports_speed": true,
+        "supports_fast_mode": true,
         "prompt_cache_min_tokens": 512,
         "source": "https://platform.claude.com/docs/en/about-claude/pricing"
     },
@@ -14103,6 +14104,7 @@
         },
         "supports_output_config": true,
         "supports_speed": true,
+        "supports_fast_mode": true,
         "prompt_cache_min_tokens": 1024,
         "source": "https://platform.claude.com/docs/en/about-claude/pricing"
     },

model_prices_and_context_window.schema.json

@@ -716,6 +716,9 @@
         "supports_embedding_image_input": {
           "type": "boolean"
         },
+        "supports_fast_mode": {
+          "type": "boolean"
+        },
         "supports_forced_tool_use": {
           "type": "boolean"
         },

tests/test_litellm/test_router.py

@@ -12206,6 +12206,83 @@ def test_model_group_info_reasoning_efforts_are_unknown_when_any_deployment_is_o
 
 
 
+@pytest.mark.parametrize(
+    "model,provider,expected",
+    [
+        ("anthropic/claude-opus-5", None, True),
+        ("claude-opus-4-8", None, True),
+        ("anthropic/claude-opus-4-7", None, False),
+        ("anthropic/claude-opus-4-6", None, False),
+        ("anthropic/claude-sonnet-5", None, False),
+        ("anthropic/off-map-opus", None, False),
+        ("vertex_ai/claude-opus-5", None, False),
+        ("bedrock/claude-opus-5", None, False),
+        ("claude-opus-5", "vertex_ai", False),
+        ("claude-opus-5", "bedrock", False),
+    ],
+)
+@pytest.mark.parametrize("operator_flag", [True, False])
+def test_model_group_info_fast_mode_uses_exact_provider_catalog(
+    local_model_cost_map: None, model: str, provider: str | None, expected: bool, operator_flag: bool
+) -> None:
+    router: Final = Router(model_list=[{
+        "model_name": "fast-group",
+        "litellm_params": {"model": model, "custom_llm_provider": provider, "api_key": "fake-key"},
+        "model_info": {"supports_fast_mode": operator_flag},
+    }])
+
+    result: Final = router.get_model_group_info("fast-group")
+
+    assert result is not None
+    assert result.supports_fast_mode is expected
+
+
+@pytest.mark.parametrize("flag", [None, False, "true", 1])
+def test_model_group_info_fast_mode_fails_closed_without_explicit_boolean(
+    local_model_cost_map: None, monkeypatch: pytest.MonkeyPatch, flag: object
+) -> None:
+    entry: Final = {key: value for key, value in litellm.model_cost["claude-opus-5"].items()
+                   if key != "supports_fast_mode"}
+    if flag is not None:
+        entry["supports_fast_mode"] = flag
+    monkeypatch.setitem(litellm.model_cost, "claude-opus-5", entry)
+    router: Final = Router(model_list=[{
+        "model_name": "fast-group",
+        "litellm_params": {"model": "anthropic/claude-opus-5", "api_key": "fake-key"},
+        "model_info": {"supports_fast_mode": True},
+    }])
+
+    result: Final = router.get_model_group_info("fast-group")
+
+    assert result is not None
+    assert result.supports_fast_mode is False
+
+
+@pytest.mark.parametrize("other_model,expected", [
+    ("anthropic/claude-opus-4-8", True),
+    ("anthropic/claude-opus-4-7", False),
+    ("anthropic/off-map-opus", False),
+    ("vertex_ai/claude-opus-5", False),
+    ("bedrock/claude-opus-5", False),
+])
+@pytest.mark.parametrize("reverse", [True, False])
+def test_model_group_info_fast_mode_requires_every_deployment(
+    local_model_cost_map: None, other_model: str, expected: bool, reverse: bool
+) -> None:
+    models: Final = (other_model, "anthropic/claude-opus-5") if reverse else (
+        "anthropic/claude-opus-5", other_model
+    )
+    router: Final = Router(model_list=[{
+        "model_name": "fast-group",
+        "litellm_params": {"model": model, "api_key": "fake-key"},
+    } for model in models])
+
+    result: Final = router.get_model_group_info("fast-group")
+
+    assert result is not None
+    assert result.supports_fast_mode is expected
+
+
 def test_model_group_info_surfaces_supports_parallel_function_calling(local_model_cost_map):
     """``/model_group/info`` folds each deployment's registry flags into the group; a deployment whose
     registry entry declares parallel function calling must flip the group to True instead of False."""

tests/test_litellm/test_utils.py

@@ -1103,6 +1103,7 @@ def test_aaamodel_prices_and_context_window_json_is_valid():
                 "supports_sampling_params": {"type": "boolean"},
                 "supports_output_config": {"type": "boolean"},
                 "supports_speed": {"type": "boolean"},
+                "supports_fast_mode": {"type": "boolean"},
                 "supported_audio_formats": {
                     "type": "array",
                     "items": {

ui/litellm-dashboard/src/components/add_model/ComplexityRouterConfig.tsx

@@ -42,9 +42,10 @@ import { Restricted, restrictedBy } from "./TierRestrictions";
 import { type TierSetAction, applyTierSetAction, setFallbackTier } from "./tier_set_actions";
 import {
   ReasoningEffort,
+  TierModelParamChange,
   TierModelParamsByTier,
   classifierEffortOptionsForModels,
-  setTierModelReasoningEffort,
+  setTierModelParam,
   tierEffortOptionsForModels,
   tierRowLabel,
 } from "./complexity_router_tiers";
@@ -621,6 +622,9 @@ const ComplexityRouterConfig: React.FC<ComplexityRouterConfigProps> = ({
   const exitToBuiltInTiers = () => dispatch({ kind: "restore" });
 
   const tierEffortOptionsByModel = tierEffortOptionsForModels(modelInfo);
+  const fastModeByModel = Object.fromEntries(
+    modelInfo.map((model) => [model.model_group, model.supports_fast_mode === true]),
+  );
   const classifierEffortOptionsByModel = classifierEffortOptionsForModels(modelInfo);
 
   // Embedding models can't serve a chat-completion role, so they're excluded here.
@@ -631,12 +635,11 @@ const ComplexityRouterConfig: React.FC<ComplexityRouterConfigProps> = ({
       label: model.model_group,
     }));
 
-  const handleTierModelEffortChange = (tier: string, model: string, effort: ReasoningEffort | undefined) => {
+  const handleTierModelParamChange = (tier: string, model: string, change: TierModelParamChange) =>
     onChange({
       ...value,
-      tier_model_params: setTierModelReasoningEffort(value.tier_model_params, tier, model, effort),
+      tier_model_params: setTierModelParam(value.tier_model_params, tier, model, change),
     });
-  };
 
   // Clearing the select drops the key entirely rather than storing "", so an emptied pin reads as
   // "track the tiers" everywhere downstream instead of as a blank model name.
@@ -734,7 +737,13 @@ const ComplexityRouterConfig: React.FC<ComplexityRouterConfigProps> = ({
                     models={row.models}
                     effortOptionsByModel={tierEffortOptionsByModel}
                     paramsByModel={row.params}
-                    onEffortChange={(model, effort) => handleTierModelEffortChange(row.id, model, effort)}
+                    fastModeByModel={fastModeByModel}
+                    onEffortChange={(model, effort) =>
+                      handleTierModelParamChange(row.id, model, ["reasoning_effort", effort])
+                    }
+                    onFastModeChange={(model, enabled) =>
+                      handleTierModelParamChange(row.id, model, ["speed", enabled ? "fast" : undefined])
+                    }
                   />
                   {row.models.length > 1 && (
                     <span className="text-xs text-muted-foreground">

ui/litellm-dashboard/src/components/add_model/ComplexityRouterFastMode.integration.test.tsx

@@ -0,0 +1,143 @@
+import userEvent from "@testing-library/user-event";
+import React from "react";
+import { describe, expect, it, vi } from "vitest";
+import { renderWithProviders, screen } from "../../../tests/test-utils";
+import {
+  buildUpdatedComplexityRouterConfig,
+  hydrateComplexityRouterConfig,
+} from "../edit_auto_router/edit_auto_router_modal";
+import type { ModelGroup } from "../llm_calls/fetch_models";
+import ComplexityRouterConfig, { type ComplexityRouterConfigValue } from "./ComplexityRouterConfig";
+
+const modelInfo: ModelGroup[] = [
+  { model_group: "primary", supported_reasoning_efforts: ["low", "high"], supports_fast_mode: true },
+  { model_group: "secondary", supports_fast_mode: true },
+  { model_group: "blocked", supported_reasoning_efforts: ["low"], supports_fast_mode: false },
+  { model_group: "missing", supported_reasoning_efforts: ["low"] },
+];
+
+it.each([false, true])("edits and round-trips independent model settings with custom tiers=%s", async (custom) => {
+  const user = userEvent.setup();
+  const tier = custom ? "custom-a" : "COMPLEX";
+  const otherTier = custom ? "custom-b" : "REASONING";
+  const label = custom ? "Interactive" : "Complex";
+  const models = ["primary", "secondary", "blocked", "missing"];
+  const initial: ComplexityRouterConfigValue = {
+    tiers: { SIMPLE: [], MEDIUM: [], COMPLEX: models, REASONING: ["primary"] },
+    classifier_type: "heuristic",
+    ...(custom && {
+      custom_tier_set: {
+        tiers: [
+          { id: tier, name: label, definition: "Interactive requests", models },
+          { id: otherTier, name: "Deliberate", definition: "Careful requests", models: ["primary"] },
+        ],
+        fallback_tier_id: tier,
+      },
+    }),
+    tier_model_params: {
+      [tier]: {
+        primary: { reasoning_effort: "high", max_tokens: 1024 },
+        secondary: { speed: "fast" },
+        blocked: { speed: "fast" },
+      },
+      [otherTier]: { primary: { speed: "fast", reasoning_effort: "low" } },
+    },
+  };
+  const onChange = vi.fn<(value: ComplexityRouterConfigValue) => void>();
+  const editor = (value: ComplexityRouterConfigValue) => (
+    <ComplexityRouterConfig modelInfo={modelInfo} value={value} onChange={onChange} />
+  );
+  const view = renderWithProviders(editor(initial));
+  const fast = () => screen.getByRole("switch", { name: `Fast mode for primary in the ${label} tier` });
+
+  expect(screen.getAllByRole("switch", { name: /^Fast mode for/ })).toHaveLength(3);
+  expect(screen.queryByRole("switch", { name: /^Fast mode for (blocked|missing)/ })).not.toBeInTheDocument();
+  expect(screen.queryByRole("combobox", { name: /^Reasoning effort for secondary/ })).not.toBeInTheDocument();
+  expect(screen.getByRole("switch", { name: `Fast mode for secondary in the ${label} tier` })).toBeChecked();
+  expect(fast()).not.toBeChecked();
+  expect(onChange).not.toHaveBeenCalled();
+
+  await user.click(fast());
+  const enabled = onChange.mock.lastCall![0];
+  expect(enabled.tier_model_params).toEqual({
+    ...initial.tier_model_params,
+    [tier]: {
+      ...initial.tier_model_params![tier],
+      primary: { reasoning_effort: "high", max_tokens: 1024, speed: "fast" },
+    },
+  });
+  const saved = buildUpdatedComplexityRouterConfig({}, enabled);
+  expect(saved.tier_model_configs).toEqual({
+    [custom ? label : tier]: [
+      { model_name: "primary", litellm_params: { reasoning_effort: "high", max_tokens: 1024, speed: "fast" } },
+      { model_name: "secondary", litellm_params: { speed: "fast" } },
+      { model_name: "blocked", litellm_params: { speed: "fast" } },
+    ],
+    [custom ? "Deliberate" : otherTier]: [
+      { model_name: "primary", litellm_params: { speed: "fast", reasoning_effort: "low" } },
+    ],
+  });
+  const reopened = hydrateComplexityRouterConfig(saved, undefined);
+  const reopenedTier = custom ? reopened.custom_tier_set!.tiers[0].id : tier;
+  view.rerender(editor(reopened));
+  expect(fast()).toBeChecked();
+
+  await user.click(screen.getByRole("combobox", { name: `Reasoning effort for primary in the ${label} tier` }));
+  await user.click(await screen.findByRole("option", { name: "low" }));
+  const effortChanged = onChange.mock.lastCall![0];
+  expect(effortChanged.tier_model_params?.[reopenedTier].primary).toEqual({
+    reasoning_effort: "low",
+    max_tokens: 1024,
+    speed: "fast",
+  });
+  view.rerender(editor(effortChanged));
+  await user.click(fast());
+  const disabled = onChange.mock.lastCall![0];
+  expect(disabled.tier_model_params).toEqual({
+    ...effortChanged.tier_model_params,
+    [reopenedTier]: {
+      ...effortChanged.tier_model_params![reopenedTier],
+      primary: { reasoning_effort: "low", max_tokens: 1024 },
+    },
+  });
+  view.rerender(editor(disabled));
+  expect(fast()).not.toBeChecked();
+
+  const picker = () => screen.getByRole("combobox", { name: `Select model(s) for ${label.toLowerCase()} queries` });
+  await user.click(picker());
+  await user.click(await screen.findByRole("option", { name: "primary" }));
+  await user.keyboard("{Escape}");
+  const deselected = onChange.mock.lastCall![0];
+  expect(deselected.tier_model_params?.[reopenedTier]).toEqual({
+    secondary: { speed: "fast" },
+    blocked: { speed: "fast" },
+  });
+  view.rerender(editor(deselected));
+  expect(screen.queryByRole("switch", { name: `Fast mode for primary in the ${label} tier` })).not.toBeInTheDocument();
+  await user.click(picker());
+  await user.click(await screen.findByRole("option", { name: "primary" }));
+  await user.keyboard("{Escape}");
+  const reselected = onChange.mock.lastCall![0];
+  view.rerender(editor(reselected));
+  expect(fast()).not.toBeChecked();
+  expect(screen.getByRole("combobox", { name: `Reasoning effort for primary in the ${label} tier` })).toHaveTextContent(
+    "Default",
+  );
+});
+
+describe("Fast mode metadata", () => {
+  it("offers nothing before model capabilities load and leaves stored speed untouched", () => {
+    const value: ComplexityRouterConfigValue = {
+      tiers: { SIMPLE: ["primary"], MEDIUM: [], COMPLEX: [], REASONING: [] },
+      classifier_type: "heuristic",
+      tier_model_params: { SIMPLE: { primary: { speed: "fast" } } },
+    };
+    const onChange = vi.fn();
+    renderWithProviders(<ComplexityRouterConfig modelInfo={[]} value={value} onChange={onChange} />);
+    expect(screen.queryByRole("switch", { name: /^Fast mode for/ })).not.toBeInTheDocument();
+    expect(onChange).not.toHaveBeenCalled();
+    expect(buildUpdatedComplexityRouterConfig({}, value).tier_model_configs).toEqual({
+      SIMPLE: [{ model_name: "primary", litellm_params: { speed: "fast" } }],
+    });
+  });
+});

ui/litellm-dashboard/src/components/add_model/TierModelEffortRows.tsx

@@ -1,5 +1,6 @@
 import { Select, SelectContent, SelectItem, SelectTrigger, SelectValue } from "@/components/ui/select";
 import { SimpleTooltip } from "@/components/ui/tooltip";
+import { Switch } from "@/components/ui/switch";
 import { Info } from "lucide-react";
 import React from "react";
 import { ReasoningEffort, TierModelParams } from "./complexity_router_tiers";
@@ -18,6 +19,8 @@ interface TierModelEffortRowsProps {
   effortOptionsByModel: Record<string, string[]>;
   paramsByModel: Record<string, TierModelParams> | undefined;
   onEffortChange: (model: string, effort: ReasoningEffort | undefined) => void;
+  fastModeByModel?: Record<string, boolean>;
+  onFastModeChange: (model: string, enabled: boolean) => void;
 }
 
 export interface TierEffortRow {
@@ -29,13 +32,16 @@ export interface TierEffortRow {
 /**
  * A stored effort outside the model's supported set (hand-authored, or capabilities changed since
  * it was saved) is listed anyway, so the row renders with its value selected and can be cleared.
- * Only a model with no supported level and nothing stored drops out.
  */
 export const tierEffortRows = ({
   models,
   effortOptionsByModel,
   paramsByModel,
-}: Pick<TierModelEffortRowsProps, "models" | "effortOptionsByModel" | "paramsByModel">): TierEffortRow[] =>
+  fastModeByModel,
+}: Pick<
+  TierModelEffortRowsProps,
+  "models" | "effortOptionsByModel" | "paramsByModel" | "fastModeByModel"
+>): TierEffortRow[] =>
   models
     .map((model) => {
       const effort = storedEffort(paramsByModel?.[model]);
@@ -43,56 +49,74 @@ export const tierEffortRows = ({
       const listed = effort !== undefined && !supported.includes(effort) ? [...supported, effort] : supported;
       return { model, effort, options: Array.from(new Set(listed)) };
     })
-    .filter(({ options }) => options.length > 0);
+    .filter(({ model, options }) => options.length > 0 || fastModeByModel?.[model] === true);
 
-const TierModelEffortRows: React.FC<TierModelEffortRowsProps> = ({
-  tierLabel,
-  models,
-  effortOptionsByModel,
-  paramsByModel,
-  onEffortChange,
-}) => {
-  const rows = tierEffortRows({ models, effortOptionsByModel, paramsByModel });
+const TierModelEffortRows: React.FC<TierModelEffortRowsProps> = (props) => {
+  const { tierLabel, paramsByModel, onEffortChange, fastModeByModel, onFastModeChange } = props;
+  const rows = tierEffortRows(props);
   if (rows.length === 0) return null;
   return (
     <div className="mt-2 space-y-1">
-      <div className="flex items-center gap-1">
-        <span className="text-xs font-medium text-muted-foreground">Reasoning effort</span>
-        <SimpleTooltip
-          content={`Sent as reasoning_effort on requests this tier routes to the model, overriding the caller's value. Default leaves the request untouched.`}
-        >
-          <Info className="size-3 text-muted-foreground/70" />
-        </SimpleTooltip>
-      </div>
-      {rows.map(({ model, effort, options }) => (
-        <div key={model} className="flex items-center justify-between gap-2">
-          <span className="truncate text-xs">{model}</span>
-          <Select
-            items={[
-              { value: PROVIDER_DEFAULT, label: "Default" },
-              ...options.map((option) => ({ value: option, label: option })),
-            ]}
-            value={effort ?? PROVIDER_DEFAULT}
-            onValueChange={(selected: string | null) =>
-              selected !== null && onEffortChange(model, selected === PROVIDER_DEFAULT ? undefined : selected)
-            }
+      {rows.some(({ options }) => options.length > 0) && (
+        <div className="flex items-center gap-1">
+          <span className="text-xs font-medium text-muted-foreground">Reasoning effort</span>
+          <SimpleTooltip
+            content={`Sent as reasoning_effort on requests this tier routes to the model, overriding the caller's value. Default leaves the request untouched.`}
           >
-            <SelectTrigger
-              size="sm"
-              className="w-36"
-              aria-label={`Reasoning effort for ${model} in the ${tierLabel} tier`}
-            >
-              <SelectValue />
-            </SelectTrigger>
-            <SelectContent>
-              <SelectItem value={PROVIDER_DEFAULT}>Default</SelectItem>
-              {options.map((option) => (
-                <SelectItem key={option} value={option}>
-                  {option}
-                </SelectItem>
-              ))}
-            </SelectContent>
-          </Select>
+            <Info className="size-3 text-muted-foreground/70" />
+          </SimpleTooltip>
+        </div>
+      )}
+      {rows.map(({ model, effort, options }) => (
+        <div key={model} className="flex flex-wrap items-center justify-between gap-2">
+          <span className="min-w-0 flex-1 basis-32 truncate text-xs" title={model}>
+            {model}
+          </span>
+          <div className="flex flex-wrap items-center gap-3">
+            {options.length > 0 && (
+              <Select
+                items={[
+                  { value: PROVIDER_DEFAULT, label: "Default" },
+                  ...options.map((option) => ({ value: option, label: option })),
+                ]}
+                value={effort ?? PROVIDER_DEFAULT}
+                onValueChange={(selected: string | null) =>
+                  selected !== null && onEffortChange(model, selected === PROVIDER_DEFAULT ? undefined : selected)
+                }
+              >
+                <SelectTrigger
+                  size="sm"
+                  className="w-36"
+                  aria-label={`Reasoning effort for ${model} in the ${tierLabel} tier`}
+                >
+                  <SelectValue />
+                </SelectTrigger>
+                <SelectContent>
+                  <SelectItem value={PROVIDER_DEFAULT}>Default</SelectItem>
+                  {options.map((option) => (
+                    <SelectItem key={option} value={option}>
+                      {option}
+                    </SelectItem>
+                  ))}
+                </SelectContent>
+              </Select>
+            )}
+            {fastModeByModel?.[model] === true && (
+              <SimpleTooltip content="Fast mode has higher pricing and requires an eligible provider account. Off removes this tier's speed override and inherits the request or provider default">
+                <label
+                  className="flex items-center gap-2 text-xs"
+                  aria-label={`Fast mode for ${model} in the ${tierLabel} tier`}
+                >
+                  <Switch
+                    size="sm"
+                    checked={paramsByModel?.[model]?.speed === "fast"}
+                    onCheckedChange={(enabled) => onFastModeChange(model, enabled)}
+                  />
+                  Fast mode
+                </label>
+              </SimpleTooltip>
+            )}
+          </div>
         </div>
       ))}
     </div>

ui/litellm-dashboard/src/components/add_model/build_complexity_router_config.test.ts

@@ -48,6 +48,19 @@ const baseParams: BuildComplexityRouterConfigParams = {
 };
 
 describe("buildComplexityRouterConfig", () => {
+  it("carries Fast and reasoning overrides independently into a new router payload", () => {
+    const params = { speed: "fast", reasoning_effort: "high", max_tokens: 1024 };
+    const config = buildComplexityRouterConfig({
+      ...baseParams,
+      tiers: { ...tiers, COMPLEX: ["primary"], REASONING: ["secondary"] },
+      tierModelParams: { COMPLEX: { primary: params }, REASONING: { secondary: { speed: "fast" } } },
+    });
+    expect(config.tier_model_configs).toEqual({
+      COMPLEX: [{ model_name: "primary", litellm_params: params }],
+      REASONING: [{ model_name: "secondary", litellm_params: { speed: "fast" } }],
+    });
+  });
+
   it("emits tiers, classifier_type, and escalation_keywords when nothing else is configured", () => {
     const config = buildComplexityRouterConfig(baseParams);
     const expected = {

ui/litellm-dashboard/src/components/add_model/complexity_router_tiers.test.ts

@@ -7,6 +7,7 @@ import {
   serializeTierModelConfigs,
   tierRowLabel,
   setTierModelReasoningEffort,
+  setTierModelParam,
 } from "./complexity_router_tiers";
 import { resolveComplexityDefaultModel } from "./tier_rows";
 
@@ -218,6 +219,28 @@ describe("setTierModelReasoningEffort", () => {
   });
 });
 
+describe("setTierModelParam", () => {
+  it.each(["reasoning_effort", "speed"] as const)("clears only %s and preserves the input", (key) => {
+    const params = { reasoning_effort: "high", speed: "fast", max_tokens: 512 };
+    const current = { COMPLEX: { primary: params, secondary: { speed: "fast" } }, REASONING: { primary: params } };
+    const cleared = setTierModelParam(current, "COMPLEX", "primary", [key, undefined]);
+    expect(cleared).toEqual({
+      ...current,
+      COMPLEX: {
+        ...current.COMPLEX,
+        primary: key === "speed" ? { reasoning_effort: "high", max_tokens: 512 } : { speed: "fast", max_tokens: 512 },
+      },
+    });
+    expect(current.COMPLEX.primary).toEqual({ reasoning_effort: "high", speed: "fast", max_tokens: 512 });
+  });
+
+  it("removes empty records when the only override is Fast", () => {
+    const enabled = setTierModelParam(undefined, "COMPLEX", "primary", ["speed", "fast"]);
+    expect(enabled).toEqual({ COMPLEX: { primary: { speed: "fast" } } });
+    expect(setTierModelParam(enabled, "COMPLEX", "primary", ["speed", undefined])).toBeUndefined();
+  });
+});
+
 describe("pruneTierModelParams", () => {
   it("drops params for models deselected from the tier", () => {
     expect(

ui/litellm-dashboard/src/components/add_model/complexity_router_tiers.ts

@@ -114,14 +114,16 @@ export const serializeTierModelConfigs = (
   return serialized.length > 0 ? Object.fromEntries(serialized) : undefined;
 };
 
-export const setTierModelReasoningEffort = (
+export type TierModelParamChange = ["reasoning_effort", ReasoningEffort | undefined] | ["speed", "fast" | undefined];
+
+export const setTierModelParam = (
   current: TierModelParamsByTier | undefined,
   tier: string,
   model: string,
-  effort: ReasoningEffort | undefined,
+  [key, value]: TierModelParamChange,
 ): TierModelParamsByTier | undefined => {
-  const { reasoning_effort: _dropped, ...rest } = current?.[tier]?.[model] ?? {};
-  const params = effort === undefined ? rest : { ...rest, reasoning_effort: effort };
+  const { [key]: _dropped, ...rest } = current?.[tier]?.[model] ?? {};
+  const params = value === undefined ? rest : { ...rest, [key]: value };
   const byModel = Object.fromEntries(
     Object.entries({ ...current?.[tier], [model]: params }).filter(([, value]) => Object.keys(value).length > 0),
   );
@@ -131,6 +133,13 @@ export const setTierModelReasoningEffort = (
   return Object.keys(next).length > 0 ? next : undefined;
 };
 
+export const setTierModelReasoningEffort = (
+  current: TierModelParamsByTier | undefined,
+  tier: string,
+  model: string,
+  effort: ReasoningEffort | undefined,
+): TierModelParamsByTier | undefined => setTierModelParam(current, tier, model, ["reasoning_effort", effort]);
+
 export const pruneTierModelParams = (
   current: TierModelParamsByTier | undefined,
   tier: string,

ui/litellm-dashboard/src/components/llm_calls/fetch_models.test.tsx

@@ -52,6 +52,23 @@ describe("fetchAvailableModels", () => {
     ]);
   });
 
+  it("carries only explicitly supported Fast capabilities, not accepted speed parameters", async () => {
+    modelHubCallMock.mockResolvedValue({
+      data: [
+        { model_group: "fast", supports_fast_mode: true },
+        { model_group: "blocked", supports_fast_mode: false },
+        { model_group: "missing", supports_speed: true },
+        { model_group: "unknown", supports_fast_mode: null },
+      ],
+    });
+    expect(await fetchAvailableModels("token")).toEqual([
+      { model_group: "blocked" },
+      { model_group: "fast", supports_fast_mode: true },
+      { model_group: "missing" },
+      { model_group: "unknown" },
+    ]);
+  });
+
   it("preserves absent, unknown, empty, and explicit effort capability states", async () => {
     modelHubCallMock.mockResolvedValue({
       data: [

ui/litellm-dashboard/src/components/llm_calls/fetch_models.tsx

@@ -7,6 +7,7 @@ export interface ModelGroup {
   model_group: string;
   mode?: string;
   supports_reasoning?: boolean;
+  supports_fast_mode?: boolean;
   supported_reasoning_efforts?: string[] | null;
 }
 
@@ -16,6 +17,7 @@ interface AvailableModel {
   id?: string | null;
   mode?: string | null;
   supports_reasoning?: boolean | null;
+  supports_fast_mode?: boolean | null;
   supported_reasoning_efforts?: string[] | null;
 }
 
@@ -25,6 +27,7 @@ const toModelGroup = (item: AvailableModel): ModelGroup => {
     model_group: groupName,
     ...(item.mode && { mode: item.mode }),
     ...(item.supports_reasoning === true && { supports_reasoning: true }),
+    ...(item.supports_fast_mode === true && { supports_fast_mode: true }),
     ...(item.supported_reasoning_efforts !== undefined && {
       supported_reasoning_efforts: item.supported_reasoning_efforts,
     }),

ui/litellm-dashboard/src/lib/http/schema.d.ts

@@ -32708,6 +32708,11 @@ export interface components {
             supported_openai_params: string[] | null;
             /** Supported Reasoning Efforts */
             supported_reasoning_efforts?: string[] | null;
+            /**
+             * Supports Fast Mode
+             * @default false
+             */
+            supports_fast_mode: boolean;
             /**
              * Supports Function Calling
              * @default false