---

license: apache-2.0
base_model: Qwen/Qwen3.6-35B-A3B
pipeline_tag: image-text-to-text
library_name: mlx
tags:
  - mlx
  - mlx-vlm
  - apple-silicon
  - qwen3.6
  - moe
  - quantized
  - 6-bit
  - mtp
  - ax-engine
  - automatosx
---

# AX Qwen3.6 35B-A3B MLX 6-bit MTP

> **Parameter count:** approximately 35.11B logical target parameters (35B
> total, approximately 3B active per token). `6-bit` is the target quantization
> precision, not a 6B model-size claim. The separately packaged MTP sidecar is
> not included in the target count.

This is an **MLX model for Apple Silicon**. It packages the 6-bit MLX
quantization of Qwen3.6-35B-A3B together with an AX Engine-compatible
multi-token-prediction (MTP) sidecar.

This repository does not contain PyTorch or GGUF weights.

## Model details

- Base model: [Qwen/Qwen3.6-35B-A3B](https://huggingface.co/Qwen/Qwen3.6-35B-A3B)
- MLX base: [mlx-community/Qwen3.6-35B-A3B-6bit](https://huggingface.co/mlx-community/Qwen3.6-35B-A3B-6bit)
- Format: MLX Safetensors
- Architecture: mixture of experts (MoE), with 256 experts and 8 selected per token
- Quantization: 6-bit affine, group size 64, with 8-bit MoE gate overrides
- Task: image-text-to-text and conversational generation
- Context length: 262,144 tokens in the model configuration
- MTP sidecar: 20 tensors, draft depth up to 1
- Intended hardware: Apple Silicon

The base MLX conversion was produced with `mlx-vlm` 0.4.4. The MTP tensors
come from the official Qwen checkpoint and were normalized with AX Engine's
`prepare_mtp_sidecar.py` packaging flow.

## Download

```bash
hf download AutomatosX/AX-Qwen3.6-35B-A3B-MLX-6bit-MTP \
  --local-dir ./AX-Qwen3.6-35B-A3B-MLX-6bit-MTP
```

The repository is approximately 30.8 GB. Ensure that the destination has
enough free space before downloading.

## Use as an MLX model

For standard MLX-VLM inference without AX MTP acceleration:

```bash
pip install -U mlx-vlm

python -m mlx_vlm.generate \
  --model AutomatosX/AX-Qwen3.6-35B-A3B-MLX-6bit-MTP \
  --max-tokens 100 \
  --temperature 0.0 \
  --prompt "Describe this image." \
  --image <path-to-image>
```

The six `model-*.safetensors` shards are the MLX base weights. Standard
MLX-VLM loading uses those shards through `model.safetensors.index.json` and
does not use the AX-specific `mtp.safetensors` sidecar.

## Use with AX Engine MTP

Install [AX Engine](https://github.com/defai-digital/ax-engine), download the
repository, and serve the local directory:

```bash
hf download AutomatosX/AX-Qwen3.6-35B-A3B-MLX-6bit-MTP \
  --local-dir ./AX-Qwen3.6-35B-A3B-MLX-6bit-MTP

ax-engine serve ./AX-Qwen3.6-35B-A3B-MLX-6bit-MTP --port 31418
```

The MTP package includes:

- `mtp.safetensors`: the normalized MTP head.
- `mtplx_runtime.json`: draft-depth and sampler guidance.
- `ax_mtp_sidecar_manifest.json`: source hashes and transformation provenance.
- `config.json`: the MLX base configuration with MTP sidecar fields.

## Validation and provenance

The sidecar preparation gate recorded:

- Architecture: `qwen-moe-packed`
- MTP tensor count: 20
- MoE expert unpacking: enabled
- Norm transform: add 1 to MTP RMSNorm delta weights
- Exactness baseline: maximum absolute difference `0.0` at context length 2048
- Validation platform: macOS on arm64
- Validation timestamp: 2026-06-22

The provenance manifest uses repository-relative paths and public Hugging Face
URIs. It intentionally excludes local filesystem paths.

## License

This repository follows the upstream Apache License 2.0. See `LICENSE` and the
original Qwen model card for the model's terms, limitations, and responsible-use
guidance.