Support dense Qwen3.5 model architecture (qwen3_5)

[jsirish]

Description

The dense Qwen3.5 architecture (model_type: qwen3_5) is not currently supported by AF-M. Only the MoE variant (qwen3_5_moe) works.

Steps to Reproduce

afm mlx -m mlx-community/Qwen3.5-27B-8bit -p 8001

Error

Error starting MLX server: loadFailed("mlx-community/Qwen3.5-27B-8bit: Unsupported model type: qwen3_5")

Expected Behavior

AF-M should load dense Qwen3.5 models (27B, etc.) the same way it loads the MoE variants (122B-A10B, 35B-A3B).

Environment

• AF-M version: v0.9.5-cd2941e (afm-next via Homebrew)
• macOS 26 on Mac Studio M3 Ultra
• Model: mlx-community/Qwen3.5-27B-8bit (config.json has model_type: qwen3_5, architectures: ['Qwen3_5ForConditionalGeneration'])

Notes

The MoE variants work perfectly — Qwen3.5-122B-A10B-4bit and Qwen3.5-35B-A3B-4bit both load and run well. The Scripts/patches/ directory has Qwen3_5MoE.swift but no corresponding Qwen3_5.swift for the dense architecture.

[scouzi1966]

It is already implemented in afm-next (nightly build).

Note: load with --vlm option if you want vision (will run slower)

https://github.com/scouzi1966/maclocal-api/releases/tag/nightly-20260226-cd2941e

Next release will have it

brew unlink afm && brew install scouzi1966/afm/afm-next # switch to nightly
brew unlink afm-next && brew link afm # switch back to stable
ASSUMES you did a brew install scouzi1966/afm/afm previously

[scouzi1966]

I'm actually achieving 30+ toks/sec on an M1 MacBook Pro which is amazing. Released in 2022! Before even ChatGpt came into existance.

afm mlx -m mlx-community/Qwen3.5-35B-A3B-4bit -w

The -w gives you an instant Webui for chat

[scouzi1966]

Apologies. I didn't carefully read your message. Let me test the other models later on tonight.

[jsirish]

I appreciate it! Let me know if I can help test anything, I'm working on an M3 Ultra Studio with 256GB and 60 cores.

[scouzi1966]

It's now in the main branch. I'll try to build a nightly tonight. But I can't promise it.