We’ve published the Interim Labs Qwen3.5 series on Hugging Face. Six MLX 4-bit packages for local text generation on Apple Silicon: 2B, 4B and 9B, each in original Qwen3.5 and Huihui abliterated variants. Each also has a GGUF Q4_K_M counterpart.
These are standalone releases from Interim Labs. We’ve packaged and documented derivatives of existing models; the credit for the upstream models belongs to Qwen and Huihui.
Pick a size
| Size | Original MLX | Huihui abliterated MLX | Approx. weights |
|---|---|---|---|
| 2B | Model card | Model card | 1.06 GB |
| 4B | Model card | Model card | 2.37 GB |
| 9B | Model card | Model card | 5.04 GB |
The MLX packages are text-only, use 4-bit affine quantization with group size 64, and are configured for 8,192 tokens. You’ll need an MLX runtime that supports qwen3_5_text and this quantization. Check your runtime before downloading—app compatibility hasn’t been comprehensively verified.
Credit where it’s due
Thanks to Qwen, Huihui, and the conversion sources: mlx-community, TheWirelessPhoenix, and vanch007. The Huihui 2B package was converted using MLX-VLM and MLX.
The linked model cards cover source provenance, licenses and limitations, and link all twelve MLX and GGUF repositories. We’re not claiming benchmark gains or a measured refusal-rate improvement for the abliterated variants.
