[claude] Qwen3 two-model routing via task complexity classifier (#1065) v2 (#1233)

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

config/providers.yaml

@@ -25,6 +25,19 @@ providers:
     tier: local
     url: "http://localhost:11434"
     models:
+      # ── Dual-model routing: Qwen3-8B (fast) + Qwen3-14B (quality) ──────────
+      # Both models fit simultaneously: ~6.6 GB + ~10.5 GB = ~17 GB combined.
+      # Requires OLLAMA_MAX_LOADED_MODELS=2 (set in .env) to stay hot.
+      # Ref: issue #1065 — Qwen3-8B/14B dual-model routing strategy
+      - name: qwen3:8b
+        context_window: 32768
+        capabilities: [text, tools, json, streaming, routine]
+        description: "Qwen3-8B Q6_K — fast router for routine tasks (~6.6 GB, 45-55 tok/s)"
+      - name: qwen3:14b
+        context_window: 40960
+        capabilities: [text, tools, json, streaming, complex, reasoning]
+        description: "Qwen3-14B Q5_K_M — complex reasoning and planning (~10.5 GB, 20-28 tok/s)"
+
       # Text + Tools models
       - name: qwen3:30b
         default: true
@@ -187,6 +200,20 @@ fallback_chains:
     - dolphin3          # base Dolphin 3.0 8B (uncensored, no custom system prompt)
     - qwen3:30b         # primary fallback — usually sufficient with a good system prompt
 
+  # ── Complexity-based routing chains (issue #1065) ───────────────────────
+  # Routine tasks: prefer Qwen3-8B for low latency (~45-55 tok/s)
+  routine:
+    - qwen3:8b              # Primary fast model
+    - llama3.1:8b-instruct  # Fallback fast model
+    - llama3.2:3b           # Smallest available
+
+  # Complex tasks: prefer Qwen3-14B for quality (~20-28 tok/s)
+  complex:
+    - qwen3:14b             # Primary quality model
+    - hermes4-14b           # Native tool calling, hybrid reasoning
+    - qwen3:30b             # Highest local quality
+    - qwen2.5:14b           # Additional fallback
+
 # ── Custom Models ───────────────────────────────────────────────────────────
 # Register custom model weights for per-agent assignment.
 # Supports GGUF (Ollama), safetensors, and HuggingFace checkpoint dirs.