---
license: apache-2.0
base_model:
  - WaveCut/Nanbeige4.2-3B-heretic
pipeline_tag: text-generation
library_name: mlx
tags:
  - mlx
  - dwq
  - 4-bit
  - nanbeige
  - heretic
  - coding
  - agents
---

# Nanbeige4.2-3B Heretic MLX DWQ 4-bit

A 4-bit, group-size-32 Distilled Weight Quantization (DWQ) release of
[`WaveCut/Nanbeige4.2-3B-heretic`](https://huggingface.co/WaveCut/Nanbeige4.2-3B-heretic).

This repository includes a small trusted-code MLX-LM adapter because Nanbeige
reuses 22 physical decoder layers over two loops and is not a standard Llama
layout at runtime. The adapter preserves shared weights while allocating 44
independent KV caches, one for each loop/layer execution. Unsupported optional
Nanbeige architectures are rejected explicitly.

## DWQ calibration

- 4 bits, group size 32.
- 1,024 training samples and 32 validation samples.
- Maximum sequence length: 1,025 tokens.
- Seed: 20260722.
- Corpus: 528 agentic trajectories plus 528 coding-reasoning examples,
  deterministically shuffled.
- Corpus SHA-256: `a7cfdbe02c124304bf1282bbd5ed7162bfa72dec6750b60ed2d3a68000c7a554`.
- Agentic source: `TIGER-Lab/SWE-QA-Pro-SFT-Trajectories` at
  `b8f5b8a8dcf90bca8b6d70adedac0d20dca02b86`.
- Coding source: `nvidia/OpenCodeReasoning` at
  `20a1ca19c0d050fe9057fc08339d6b370ec1c67a`.

| Validation loss | Value |
| --- | ---: |
| Initial RTN | 0.284 |
| Final DWQ | 0.043 |

MLX-LM revision: `cf10f962b7a20e63a6df43dbf0faf06070153d40`.

## Usage

The model file is repository code, so load it only after reviewing
`nanbeige_mlx.py` and pass `--trust-remote-code`.

```bash
mlx_lm.generate \
  --model WaveCut/Nanbeige4.2-3B-heretic-MLX-DWQ-4bit \
  --trust-remote-code \
  --prompt "Implement a bounded async worker pool in Python." \
  --max-tokens 256
```

Exact artifact hashes and clean-load smoke-test results are recorded in
`release-manifest.json`.