feat: add tool parser support for Qwen 3.5/3.6 (non-Coder) models

[TinyWorkshopDesign]

Feature Description

Qwen 3.5 35B A3B and Qwen 3.6 27B do not emit tool_calls when used with mlx_lm.server because the tool parser auto-detection (_infer_tool_parser) does not match their chat template patterns.

Current state

• tool_parsers/qwen3_coder.py exists but only activates for Qwen3-Coder models (chat template contains <tool_call>\n<function=)
• Regular Qwen 3.5/3.6 chat templates do NOT contain these markers
• _infer_tool_parser() returns None → no tool parsing → models respond with empty content when tools are provided

Steps to reproduce

curl -s http://localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"/Users/micky/MLX/qwen-35b-a3b-8bit","messages":[{"role":"user","content":"What is the weather in Milan?"}],"tools":[{"type":"function","function":{"name":"get_weather","description":"Get weather","parameters":{"type":"object","properties":{"city":{"type":"string"}}}}}],"max_tokens":80}'

Result: content: "" and no tool_calls in the response.

Expected behavior

Qwen 3.5/3.6 models should emit structured tool_calls via the existing tool parser infrastructure, or via a dedicated qwen3 parser.

Environment

• mlx-lm 0.31.3 (git main, 2026-05-21)
• macOS 26.5, Apple Silicon (M5 Max)
• Models: mlx-community/qwen-35b-a3b-8bit, qwen-27b-bf16

Related

• Qwen3.5-35B-A3B: Multi-turn tool calling degrades with mlx-community 4-bit/8-bit checkpoints, but not with GGUF Q4_K_XL or cloud fp #1011 (closed) — tool calling for Qwen3.5-35B-A3B was fixed but may only apply to specific checkpoints
• Improve reasoning and tool call parsing in server #711 (closed) — improved reasoning and tool call parsing

Workaround / request

Either:

1. Add Qwen 3.5/3.6 regular chat template patterns to _infer_tool_parser()
2. Allow users to specify --tool-parser qwen3_coder (or similar) explicitly via CLI flag
3. Document which models currently support tool calling and how to enable it