---
library_name: mlx
license: apache-2.0
base_model: apus-ailab/APUS-Reflex-4B
base_model_relation: quantized
pipeline_tag: text-generation
language:
- en
- zh
tags:
- apus-reflex
- decision-model
- mlx
- apple-silicon
---

# APUS Reflex 4B-MLX-4bit

> **Branding note:** This model family is now **APUS Reflex**. Some weight filenames, script/module names, served-model aliases, and the separate evaluation dataset ID retain their historical identifiers for download and runtime compatibility; these do not indicate different model versions.

[English](README.md) | [中文](README.zh-CN.md) · [Source model](https://huggingface.co/apus-ailab/APUS-Reflex-4B) · [Collection](https://huggingface.co/collections/apus-ailab/apus-reflex-6ab1ee888eb002fcdd3a2825) · [GGUF collection](https://huggingface.co/collections/apus-ailab/apus-reflex-gguf-6ab39d5e724c4d8a1021198f) · [MLX collection](https://huggingface.co/collections/apus-ailab/apus-reflex-mlx-6ab39d5fc988a1b1cb89dfc8) · [MLX-8bit](https://huggingface.co/apus-ailab/APUS-Reflex-4B-MLX-8bit) · [GGUF / Ollama](https://huggingface.co/apus-ailab/APUS-Reflex-4B-GGUF)

MLX weights (mixed 4/6-bit affine (mlx-lm `mixed_4_6`), group size 64) of [APUS-Reflex-4B](https://huggingface.co/apus-ailab/APUS-Reflex-4B) for **Apple Silicon Macs** (mlx-lm, LM Studio).

APUS Reflex is a **decision model**: each request supplies a state, an instruction and 2–16 candidates, and the model scores candidate labels A–P. It is not a chat model.

## Quick start

```bash
pip install mlx-lm
hf download apus-ailab/APUS-Reflex-4B-MLX-4bit --local-dir ./apus-reflex
python ./apus-reflex/examples/openjev_mlx.py --model ./apus-reflex
```

`examples/openjev_mlx.py` renders prompts with [openjev_contracts.py](openjev_contracts.py) (the training contract) and returns the exact candidate distribution.

## Parity

Frozen80 with identical prompt tokens, compared with the HF BF16 release (full depth, **66/80 · 82.50%**):

| Run / 运行 | Backend / 后端 | Frozen80 | = HF BF16 | Max Δp |
|---|---|---:|---:|---:|
| NVIDIA RTX PRO 6000 (CUDA) | mlx 0.32.2 on Linux x86_64 (Device(gpu, 0)) | 67/80 · 83.75% | 79/80 | 0.5980 |
| Apple M5 24 GB (Metal) | mlx 0.32.2 on Darwin arm64 (Device(gpu, 0)) | 67/80 · 83.75% | 79/80 | 0.5980 |

Verified on an Apple M5, 24 GB: peak memory 3.48 GB. The same file on CUDA and Metal gave identical decisions on 80/80 prompts (max Δp 0.043). Frozen80 is a reused development panel, not a blind benchmark. Per-question rows (candidate probabilities, choice, correctness; join with Frozen80 by `panel_index`): [evaluation/per-question/](evaluation/per-question/).

## Conversion

- mlx-lm `0.31.3` / mlx `0.32.2`; mixed 4/6-bit affine (mlx-lm `mixed_4_6`), group size 64.
- 6-bit lm_head and v_proj/down_proj in sensitive layers, 4-bit elsewhere (the MLX analogue of Q4_K_M); MoE router gates stay 8-bit ([recipe](conversion.json)). Calibrated DWQ/GPTQ were tried and were not practical for this hybrid-attention model on the conversion hardware.
- GDN `A_log` and `linear_attn.norm.weight` keep their source precision (FP32 in the 35B release).
- Full depth only, text only, probabilities **not calibrated**.

## License

Apache-2.0, inherited from the source model; see [LICENSE](LICENSE). Base model: [Qwen/Qwen3.5-4B](https://huggingface.co/Qwen/Qwen3.5-4B).

**Authors:** gumpcheng ([xDAN2099](https://huggingface.co/xDAN2099)), zhangxu, [APUS AI-LAB](https://github.com/APUS-AI-Lab).