tfjack/Qwen3.6-35B-A3B-oQ8e-fp16-mtp

Qwen3.6-35B-A3B-oQ8e-fp16-mtp

This is an oQe (Enhanced Quantization) build of `Qwen/Qwen3.6-35B-A3B`, converted and optimized for Apple Silicon using oMLX.

It is a multimodal (vision-language) model — the vision tower is preserved at float16, unchanged, so image and video understanding remain fully intact alongside the quantized language model.

[!NOTE] Card generation: This model card was drafted with the assistance of an AI coding interface. All technical content — quantization details, sampling parameters, output-length guidance, and the citation — is reproduced from the official `Qwen/Qwen3.6-35B-A3B` model card and the quantization configuration.

Quantization Details

• —Base Model: `Qwen/Qwen3.6-35B-A3B`
• —Quant Scheme: oQe enhanced — imatrix-enhanced dynamic mixed-precision with sensitivity-driven bit allocation
• —Effective Bits-per-Weight: ~8.6 bpw (oQ8e)
• —Output File Size: 39.5 GB
• —Context Length: 262,144 tokens natively; up to 1,010,000 with YaRN (RoPE) scaling
• —Non-Quantized Precision: float16 (chosen for ~20% faster prefill on M1/M2; M3 and newer default to bfloat16, which is natively supported at full speed)
• —Vision Encoder: Preserved at float32
• —MTP Support: Preserved (Multi-Token Prediction heads intact)
• —Format: MLX safetensors (compatible with oMLX, mlx-lm, and LM Studio MLX runtime)

Compatibility & Runtimes

This model works out-of-the-box on Apple Silicon with:

1. 1.oMLX Engine: Native support for MTP draft heads and SSD-tiered KV prompt caching.
2. 2.`mlx-lm` / `mlx-vlm`: Standard Apple Silicon python runtime (use mlx-vlm for image input).
3. 3.LM Studio: Load via the native MLX engine (mlx-engine).

Recommended Usage (from the Base Model Card)

Because this is a quantization of `Qwen/Qwen3.6-35B-A3B`, the base model's own recommended settings apply directly to this quant. These settings come from the base model's official Best Practices and are recommended for optimal performance:

Recommended Sampling Parameters

As recommended in the base model card: use a max output length of 32,768 tokens for most queries, and 81,920 tokens for complex math/programming benchmarks.

[!Note] The YaRN (RoPE) extension to 1,010,000 tokens requires a text_config.rope_parameters change (factor: 4.0) and is supported by frameworks including transformers, vLLM, SGLang, and ktransformers (as listed in the base model card); it is not exposed in oMLX itself, where the native 262,144-token window applies.

Citation

If you use this quantization, please cite the base model:

@misc{qwen36_35b_a3b,
    title = {{Qwen3.6-35B-A3B}: Agentic Coding Power, Now Open to All},
    url = {https://qwen.ai/blog?id=qwen3.6-35b-a3b},
    author = {{Qwen Team}},
    month = {April},
    year = {2026}
}