[claude] Add vllm-mlx as high-performance local inference backend (#1069) (#1089)

Co-authored-by: Claude (Opus 4.6) <claude@hermes.local>
Co-committed-by: Claude (Opus 4.6) <claude@hermes.local>

config/providers.yaml

@@ -67,6 +67,29 @@ providers:
         capabilities: [text, creative, streaming]
         description: "Dolphin 3.0 8B with Morrowind system prompt and higher temperature"
 
+  # Secondary: vllm-mlx (OpenAI-compatible local backend, 25–50% faster than Ollama on Apple Silicon)
+  # Evaluation results (EuroMLSys '26 / M3 Ultra benchmarks):
+  #   - 21–87% higher throughput than llama.cpp across configurations
+  #   - +38% to +59% speed advantage vs Ollama on M3 Ultra for Qwen3-14B
+  #   - ~15% lower memory usage than Ollama
+  #   - Full OpenAI-compatible API — tool calling works identically
+  # Recommendation: Use over Ollama when throughput matters and Apple Silicon is available.
+  #   Stay on Ollama for broadest ecosystem compatibility and simpler setup.
+  # To enable: start vllm-mlx server (`python -m vllm.entrypoints.openai.api_server
+  #   --model Qwen/Qwen2.5-14B-Instruct-MLX --port 8000`) then set enabled: true.
+  - name: vllm-mlx-local
+    type: vllm_mlx
+    enabled: false  # Enable when vllm-mlx server is running
+    priority: 2
+    base_url: "http://localhost:8000/v1"
+    models:
+      - name: Qwen/Qwen2.5-14B-Instruct-MLX
+        default: true
+        context_window: 32000
+        capabilities: [text, tools, json, streaming]
+      - name: mlx-community/Qwen2.5-7B-Instruct-4bit
+        context_window: 32000
+        capabilities: [text, tools, json, streaming]
 
   # Tertiary: OpenAI (if API key available)
   - name: openai-backup