---
license: apache-2.0
base_model:
  - Qwen/Qwen3-Coder-Next
  - mlx-community/Qwen3-Coder-Next-bf16
pipeline_tag: text-generation
library_name: mlx
tags:
  - mlx
  - apple-silicon
  - qwen3-next
  - coding
  - mixture-of-experts
  - quantized
  - mixed-precision
  - optiq
  - 4-bit
  - 8-bit
  - automatosx
  - ax-engine
---

# AX Qwen3 Coder Next MLX OptiQ 4-bit

> **Parameter count:** 79,674,391,296 logical parameters (approximately
> 79.67B total, ~3B active per token). `4-bit` names the low OptiQ
> tier, not a 4B model-size claim.

This is an MLX OptiQ mixed-precision release of Qwen3-Coder-Next for Apple
Silicon. AutomatosX converted it from the revision-pinned BF16 MLX source
[mlx-community/Qwen3-Coder-Next-bf16](https://huggingface.co/mlx-community/Qwen3-Coder-Next-bf16) at commit `83d523e8883faaea659705840ce3560472286d08`
using [`mlx-optiq` 0.4.2](https://mlx-optiq.com/docs/cli).

## Attribution and changes

AutomatosX did **not** train Qwen3-Coder-Next. AutomatosX produced this OptiQ
conversion with exact calibration-driven per-layer KL sensitivity:

- candidate bits: 4 and 8
- group size: 64
- target BPW: 5.0 (achieved 5.0235)
- reference: `uniform_4bit` (BF16 reference does not fit in 128 GB UMA for this
  80B-class MoE; OptiQ streams BF16 weights for probes against a uniform-4-bit
  baseline)
- calibration: 24 × 512-token sequences from the six-domain OptiQ mix (prose,
  thinking, code, agent loops, tool calling, instructions)
- allocation guards: block floor 2, low-bit run limit 3, first/last protection

No new training or benchmark leaderboard claims are made by AutomatosX.
Evaluate coding and tool-use quality on your own tasks.

## Package details

| Property | Value |
| --- | --- |
| Format | MLX Safetensors |
| Architecture | Qwen3NextForCausalLM, mixture of experts |
| Quantization | OptiQ mixed 4/8-bit, group size 64 |
| Measured allocation | 110 components at 4-bit; 431 at 8-bit |
| Target / achieved BPW | 5.0 / 5.0235 |
| Layers / experts | 48 layers · 512 experts · 10 active |
| Configured context | 262,144 tokens |
| Logical parameters | 79,674,391,296 (79.67B) |
| Download size | Approximately 55.1 GB |
| Intended hardware | Apple Silicon |

## Download

```bash
hf download AutomatosX/AX-Qwen3-Coder-Next-MLX-OptiQ-4bit \
  --local-dir ./AX-Qwen3-Coder-Next-MLX-OptiQ-4bit
```

## Use with MLX-LM

```bash
pip install -U mlx-lm
mlx_lm.generate \
  --model AutomatosX/AX-Qwen3-Coder-Next-MLX-OptiQ-4bit \
  --prompt "Write a Python function that merges two sorted lists."
```

Apply the included chat template for conversational or tool-using prompts.

## Serve with AX Engine

```bash
ax-engine doctor --mlx-model-artifacts-dir ./AX-Qwen3-Coder-Next-MLX-OptiQ-4bit
ax-engine serve ./AX-Qwen3-Coder-Next-MLX-OptiQ-4bit --port 31418
```

## Validation

Validated on macOS arm64 with mlx-optiq 0.4.2
and AX Engine 6.11.1:

- OptiQ metadata and 4/8-bit allocation: present
- Safetensors index logical parameter override: 79,674,391,296
- MLX-LM load/generation: `passed`
- AX Engine doctor: `ready`

Sample generation (truncated): `"Here's a Python function that merges two sorted lists into a single sorted list:\n\n```python\ndef merge_sorted_lists(list1, list2):\n    \"\"\"\n    Merge two sorted l"`

## Provenance

Exact conversion inputs, allocation, and weight digests are recorded in
`ax_provenance.json`, `optiq_metadata.json`, and (when present)
`optiq_sensitivity.json`. The source model card is preserved as
`UPSTREAM_README.md`.

## License

Apache License 2.0. See `LICENSE`, the
[Qwen/Qwen3-Coder-Next model card](https://huggingface.co/Qwen/Qwen3-Coder-Next), and
`UPSTREAM_README.md` for limitations and responsible-use guidance.