Qwen3.6-35B-A3B Abliterated — native MLX, nvfp4, with working MTP speculative decoding

This is a native MLX model (Apple Silicon), quantized to nvfp4, built directly from the original full-precision (bf16) safetensors weights of huihui-ai/Huihui-Qwen3.6-35B-A3B-abliterated using Ollama’s native ollama create -q nvfp4 import pipeline.

The source weights include Qwen3.6’s MTP (Multi-Token Prediction) layers, and they survived the quantization intact — this model runs with real, working speculative decoding (confirmed ~75-80% draft acceptance rate in testing), not just a plain quantized conversion.

Why this exists

Most abliterated Qwen3.6 forks floating around are either: - GGUF (llama.cpp), noticeably slower on Apple Silicon than native MLX, or - Mislabeled “-mlx” repos that are actually GGUF under the hood (check the Ollama manifest media type or the GGUF magic bytes if in doubt) As far as I could find, no one had published an abliterated Qwen3.6-35B-A3B that is genuinely native MLX and nvfp4-quantized and keeps MTP speculative decoding working, all at once.

Performance

Measured on an M3 Pro (36GB, ~150GB/s memory bandwidth): ~56 tok/s, faster than the official non-abliterated nvfp4 tag (~48 tok/s, which lacks MTP), and well ahead of GGUF abliterated conversions (~38 tok/s).

Format note

Despite the name not saying “MLX” explicitly — on Ollama, the nvfp4 quantization is always native MLX (Apple’s framework), never GGUF. This is genuinely MLX.

Attribution & License

• Base model: Qwen/Qwen3.6-35B-A3B (Alibaba Cloud, Apache 2.0)
• Abliteration: huihui-ai/Huihui-Qwen3.6-35B-A3B-abliterated (Apache 2.0)
• This repackaging: re-quantized to native MLX nvfp4 with MTP preserved, via Ollama’s safetensors import pipeline. Full license text: https://www.apache.org/licenses/LICENSE-2.0

Usage warning

This model has had its safety alignment substantially reduced via abliteration. Outputs are not filtered. Use responsibly and at your own risk — same warnings as the upstream huihui-ai model apply.