diff --git a/tests/test_model_inference.py b/tests/test_model_inference.py
new file mode 100644
index 00000000000..cafece3d694
--- /dev/null
+++ b/tests/test_model_inference.py
@@ -0,0 +1,286 @@
+"""
+Tests for model inference functionality.
+
+These tests verify that self-hosted providers can infer model capabilities
+from similar models in the registry.
+"""
+import pytest
+import litellm
+from litellm.model_inference import (
+    extract_base_model_patterns,
+    find_similar_models,
+    aggregate_capabilities,
+    infer_model_capabilities,
+    SELF_HOSTED_PROVIDERS,
+)
+
+
+class TestExtractBaseModelPatterns:
+    """Test pattern extraction from model names."""
+    
+    def test_llama_full_pattern(self):
+        patterns = extract_base_model_patterns("my-custom-llama-3.1-70b-instruct")
+        assert "llama-3.1-70b-instruct" in patterns
+        assert "llama-3.1-70b" in patterns
+        assert "llama-3.1" in patterns
+        assert "llama" in patterns
+    
+    def test_mistral_pattern(self):
+        patterns = extract_base_model_patterns("mistral-7b-instruct-v0.2")
+        assert any("mistral" in p for p in patterns)
+        assert any("7b" in p.lower() for p in patterns)
+    
+    def test_qwen_pattern(self):
+        patterns = extract_base_model_patterns("qwen-72b-chat")
+        assert any("qwen" in p for p in patterns)
+        assert any("72b" in p.lower() for p in patterns)
+    
+    def test_removes_custom_prefix(self):
+        patterns = extract_base_model_patterns("my-llama-3-8b")
+        # Should remove "my-" prefix
+        assert "llama" in patterns
+    
+    def test_case_insensitive(self):
+        patterns = extract_base_model_patterns("Llama-3.1-70B-Instruct")
+        # All patterns should be lowercase
+        assert all(p.islower() for p in patterns)
+
+
+class TestFindSimilarModels:
+    """Test finding similar models in model_cost."""
+    
+    def test_finds_llama_models(self):
+        # We know there are llama models in model_cost
+        patterns = ["llama-3", "llama"]
+        matches = find_similar_models(patterns)
+        
+        assert len(matches) > 0
+        assert all("info" in m for m in matches)
+        assert all("key" in m for m in matches)
+    
+    def test_stops_at_first_match(self):
+        # Should find matches with first pattern and stop
+        patterns = ["llama-3.1", "completely-unknown-model"]
+        matches = find_similar_models(patterns)
+        
+        # Should only match llama-3.1, not try the unknown pattern
+        assert len(matches) > 0
+        assert all("llama" in m["key"].lower() for m in matches)
+    
+    def test_skips_inferred_models(self):
+        # Add a fake inferred model
+        litellm.model_cost["test_inferred"] = {
+            "_inferred": True,
+            "max_tokens": 1000
+        }
+        
+        matches = find_similar_models(["test_inferred"])
+        
+        # Should not include inferred models
+        assert not any(m["key"] == "test_inferred" for m in matches)
+        
+        # Cleanup
+        del litellm.model_cost["test_inferred"]
+    
+    def test_no_matches_returns_empty(self):
+        patterns = ["completely-unknown-model-xyz-123"]
+        matches = find_similar_models(patterns)
+        assert len(matches) == 0
+
+
+class TestAggregateCapabilities:
+    """Test aggregation of model capabilities."""
+    
+    def test_uses_max_for_numeric_fields(self):
+        matches = [
+            {"key": "model1", "info": {"max_tokens": 4096, "max_input_tokens": 2048}},
+            {"key": "model2", "info": {"max_tokens": 8192, "max_input_tokens": 4096}},
+            {"key": "model3", "info": {"max_tokens": 2048, "max_input_tokens": 1024}},
+        ]
+        
+        result = aggregate_capabilities(matches)
+        
+        assert result["max_tokens"] == 8192  # max value
+        assert result["max_input_tokens"] == 4096  # max value
+    
+    def test_uses_any_for_boolean_fields(self):
+        matches = [
+            {"key": "model1", "info": {"supports_vision": False, "supports_function_calling": True}},
+            {"key": "model2", "info": {"supports_vision": True, "supports_function_calling": False}},
+        ]
+        
+        result = aggregate_capabilities(matches)
+        
+        # Should be True if ANY model supports it
+        assert result["supports_vision"] is True
+        assert result["supports_function_calling"] is True
+    
+    def test_zeros_out_costs(self):
+        matches = [
+            {"key": "model1", "info": {"input_cost_per_token": 0.001, "output_cost_per_token": 0.002}},
+        ]
+        
+        result = aggregate_capabilities(matches)
+        
+        assert result["input_cost_per_token"] == 0.0
+        assert result["output_cost_per_token"] == 0.0
+    
+    def test_handles_none_values(self):
+        matches = [
+            {"key": "model1", "info": {"max_tokens": None, "supports_vision": None}},
+            {"key": "model2", "info": {"max_tokens": 4096, "supports_vision": True}},
+        ]
+        
+        result = aggregate_capabilities(matches)
+        
+        # Should ignore None values
+        assert result["max_tokens"] == 4096
+        assert result["supports_vision"] is True
+    
+    def test_empty_matches_returns_empty(self):
+        result = aggregate_capabilities([])
+        assert result == {}
+
+
+class TestInferModelCapabilities:
+    """Test end-to-end model inference."""
+    
+    def test_inference_for_llama_variant(self):
+        # Test inferring a custom llama model
+        result = infer_model_capabilities(
+            model="my-custom-llama-3-70b",
+            custom_llm_provider="hosted_vllm"
+        )
+        
+        assert result is not None
+        assert result["_inferred"] is True
+        assert "_inferred_from" in result
+        assert result["litellm_provider"] == "hosted_vllm"
+        assert result["input_cost_per_token"] == 0.0
+        assert result["output_cost_per_token"] == 0.0
+        
+        # Should have inferred some capabilities
+        assert "max_tokens" in result or "max_input_tokens" in result
+    
+    def test_caches_inferred_result(self):
+        model = "test-custom-llama-for-caching"
+        provider = "hosted_vllm"
+        cache_key = f"{provider}/{model}"
+        
+        # Clear cache if exists
+        if cache_key in litellm.model_cost:
+            del litellm.model_cost[cache_key]
+        
+        # First call - should infer
+        result1 = infer_model_capabilities(model=model, custom_llm_provider=provider)
+        
+        # Check it's cached
+        assert cache_key in litellm.model_cost
+        
+        # Second call - should use cache
+        result2 = infer_model_capabilities(model=model, custom_llm_provider=provider)
+        
+        # Should return same result
+        assert result1 == result2
+        
+        # Cleanup
+        if cache_key in litellm.model_cost:
+            del litellm.model_cost[cache_key]
+    
+    def test_caches_not_found_result(self):
+        model = "completely-unknown-model-xyz-123-test"
+        provider = "hosted_vllm"
+        cache_key = f"{provider}/{model}"
+        
+        # Clear cache if exists
+        if cache_key in litellm.model_cost:
+            del litellm.model_cost[cache_key]
+        
+        # First call - should fail to infer
+        result = infer_model_capabilities(model=model, custom_llm_provider=provider)
+        
+        assert result is None
+        
+        # Check not_found marker is cached
+        assert cache_key in litellm.model_cost
+        assert litellm.model_cost[cache_key].get("_not_found") is True
+        
+        # Cleanup
+        if cache_key in litellm.model_cost:
+            del litellm.model_cost[cache_key]
+    
+    def test_returns_none_for_non_self_hosted_provider(self):
+        # Should not infer for providers like openai, anthropic, etc.
+        result = infer_model_capabilities(
+            model="some-model",
+            custom_llm_provider="openai"
+        )
+        
+        assert result is None
+    
+    def test_all_self_hosted_providers_supported(self):
+        # Verify all providers in SELF_HOSTED_PROVIDERS work
+        for provider in SELF_HOSTED_PROVIDERS:
+            # Should attempt inference (may or may not find matches)
+            result = infer_model_capabilities(
+                model="test-llama-model",
+                custom_llm_provider=provider
+            )
+            # Just checking it doesn't crash
+            assert result is None or isinstance(result, dict)
+            
+            # Cleanup cache
+            cache_key = f"{provider}/test-llama-model"
+            if cache_key in litellm.model_cost:
+                del litellm.model_cost[cache_key]
+
+
+class TestIntegrationWithGetModelInfo:
+    """Test integration with litellm.get_model_info()."""
+    
+    def test_get_model_info_uses_inference(self):
+        # This should trigger inference for a custom model name
+        try:
+            from litellm.utils import get_model_info
+            
+            # Test with a custom llama model
+            model_info = get_model_info(
+                model="my-custom-llama-3-8b",
+                custom_llm_provider="hosted_vllm"
+            )
+            
+            # Should succeed without raising "model not mapped" error
+            assert model_info is not None
+            assert hasattr(model_info, "max_tokens") or hasattr(model_info, "max_input_tokens")
+            assert model_info.input_cost_per_token == 0.0  # Self-hosted should be free
+            assert model_info.output_cost_per_token == 0.0
+            
+        except ValueError as e:
+            if "isn't mapped yet" in str(e):
+                pytest.fail("Model inference should have prevented this error")
+            raise
+        
+        finally:
+            # Cleanup
+            cache_key = "hosted_vllm/my-custom-llama-3-8b"
+            if cache_key in litellm.model_cost:
+                del litellm.model_cost[cache_key]
+    
+    def test_unknown_model_still_fails(self):
+        # Models that can't be inferred should still fail
+        from litellm.utils import get_model_info
+        
+        with pytest.raises(ValueError, match="isn't mapped yet"):
+            get_model_info(
+                model="completely-unknown-model-xyz-no-matches",
+                custom_llm_provider="hosted_vllm"
+            )
+        
+        # Cleanup
+        cache_key = "hosted_vllm/completely-unknown-model-xyz-no-matches"
+        if cache_key in litellm.model_cost:
+            del litellm.model_cost[cache_key]
+
+
+if __name__ == "__main__":
+    pytest.main([__file__, "-v"])