Qwen3.5-4B-Text-MLX-4bit

mlx-community/Qwen3.5-4B-MLX-4bit with the unused vision tower removed: 297 vision_tower.* tensors dropped, 924 language-model tensors kept, 2.9 GB → 2.2 GB download (−667 MB). Nothing else is changed.

mlx_lm's qwen3_5 loader already discards vision tensors before load (sanitize()), so this republish is purely a download/disk win — runtime behavior is identical. Verified: maximum absolute logit delta vs the original model is exactly 0.0 on a fixed prompt; greedy argmax matches.

Built for faster-context, a local repository-retrieval subagent for Apple Silicon, where this is the default serving model (zero-shot, no fine-tuning).

Use

pip install mlx-lm
mlx_lm.generate --model droyster/Qwen3.5-4B-Text-MLX-4bit --prompt "hello"

Provenance

• Base: Alibaba Qwen3.5-4B (Apache-2.0), community 4-bit MLX conversion mlx-community/Qwen3.5-4B-MLX-4bit.
• This repo: same weights, vision tower stripped with MLX (mx.load → filter vision_tower.* → mx.save_safetensors), vision_config removed from config.json. Tokenizer files unchanged.