---

license: apache-2.0
license_link: https://huggingface.co/Qwen/Qwen3.6-35B-A3B/blob/main/LICENSE
base_model:
  - mlx-community/Qwen3.6-35B-A3B-OptiQ-4bit
  - Qwen/Qwen3.6-35B-A3B
pipeline_tag: image-text-to-text
library_name: mlx
tags:
  - mlx
  - mlx-vlm
  - apple-silicon
  - qwen3.6
  - mixture-of-experts
  - vision-language-model
  - conversational
  - quantized
  - mixed-precision
  - optiq
  - 4-bit
  - 8-bit
  - speculative-decoding
  - mtp
  - ax-engine
  - automatosx
---

# AX Qwen3.6 35B-A3B MLX OptiQ 4-bit MTP

> **Parameter count:** approximately 35.11B logical target parameters (35B
> total, approximately 3B active per token). `4-bit` is the target quantization
> precision, not a 4B model-size claim. The separately packaged MTP sidecar is
> not included in the target count.

This is a self-contained MLX package for Apple Silicon. It combines the
pinned upstream OptiQ mixed-precision MoE target with an AX Engine-compatible
multi-token-prediction (MTP) sidecar.

**Try this model locally with [AX Engine](https://github.com/defai-digital/ax-engine).**

## Attribution and changes

AutomatosX did **not** train Qwen3.6-35B-A3B or create its OptiQ quantization.
The target model comes from
[mlx-community/Qwen3.6-35B-A3B-OptiQ-4bit](https://huggingface.co/mlx-community/Qwen3.6-35B-A3B-OptiQ-4bit)
at revision `70a3aa32c7feef511182bf16aa332f37e8d82014`.

AutomatosX used AX Engine's `prepare_mtp_sidecar.py` flow to extract the MTP
head from [Qwen/Qwen3.6-35B-A3B](https://huggingface.co/Qwen/Qwen3.6-35B-A3B)
at revision `995ad96eacd98c81ed38be0c5b274b04031597b0`, unpack the MoE expert tensors,
apply the required RMSNorm-delta normalization, quantize projections to 4-bit
with group size 64, patch the runtime config, and generate the AX manifests.
The upstream OptiQ card is preserved as `UPSTREAM_README.md`.

No new training or benchmark results are claimed by AutomatosX.

## Package details

| Property | Value |
| --- | --- |
| Target format | MLX Safetensors |
| Architecture | Mixture of experts; 35B total / 3B active |
| Target quantization | OptiQ mixed 4/8-bit, group size 64 |
| OptiQ allocation | 118 components at 4-bit; 392 at 8-bit |
| Achieved target BPW | 4.5062 |
| Vision tower | Bundled BF16 sidecar |
| AX MTP sidecar | 20 logical tensors; 4-bit projections |
| Maximum draft depth | 1 |
| Configured context | 262,144 tokens |
| Intended hardware | Apple Silicon |

The package retains the upstream `optiq/mtp.safetensors` file and adds the
AX-prepared root `mtp.safetensors`. AX Engine uses the root sidecar through the
`mlx_lm_extra_tensors` config entry.

## Download and serve

```bash
hf download AutomatosX/AX-Qwen3.6-35B-A3B-MLX-OptiQ-4bit-MTP \
  --local-dir ./AX-Qwen3.6-35B-A3B-MLX-OptiQ-4bit-MTP

ax-engine doctor \
  --mlx-model-artifacts-dir ./AX-Qwen3.6-35B-A3B-MLX-OptiQ-4bit-MTP

ax-engine serve ./AX-Qwen3.6-35B-A3B-MLX-OptiQ-4bit-MTP --port 31418
```

The download is approximately 25 GB. The server exposes an OpenAI-compatible
API. Consult the [AX Engine repository](https://github.com/defai-digital/ax-engine)
for installation and API examples.

## AX-specific files

- `mtp.safetensors`: AX-prepared MTP sidecar
- `mtplx_runtime.json`: draft-depth and sampler guidance
- `ax_mtp_sidecar_manifest.json`: sanitized, revision-pinned provenance
- `model-manifest.json`: AX native target manifest
- `config.json`: upstream target config with the AX sidecar registration

## Validation

Validated on macOS arm64 with AX Engine 6.9.0 on 2026-07-20:

- AX artifact doctor: `ready`, with no model issues
- Safetensors headers, data bounds, and index mappings: passed
- MTP tensor-layout exactness baseline: maximum absolute difference `0.0` at
  context length 2,048
- Source and destination revisions are immutable and recorded in the
  provenance manifest

Quantization can change model quality, and speculative-decoding acceptance
depends on the workload. Evaluate this package on your own tasks.

## License

Apache License 2.0. See `LICENSE`, the original Qwen model card, and the pinned
upstream OptiQ card for limitations and responsible-use guidance.