---
license: apache-2.0
base_model: Kwaipilot/KAT-Coder-V2.5-Dev
library_name: mlx
tags:
  - mlx
  - kat-coder
  - qwen3_5_moe
  - code
  - quantized
pipeline_tag: text-generation
---

# KAT-Coder-V2.5-Dev-MLX-4bit

MLX **4bit** (affine, group size 64) quantized
variant of [Kwaipilot/KAT-Coder-V2.5-Dev](https://huggingface.co/Kwaipilot/KAT-Coder-V2.5-Dev) for Apple
silicon via [mlx-lm](https://github.com/ml-explore/mlx-lm).

## Provenance

- Source: [Kwaipilot/KAT-Coder-V2.5-Dev](https://huggingface.co/Kwaipilot/KAT-Coder-V2.5-Dev) @ revision
  `7be56fe773e72b6f5ca93c1ae45d828ddb893922` (Apache-2.0).
- Quantized with `mlx_lm.convert` (mlx-lm 0.31.3): affine,
  4-bit, group size 64.
- **Text-only pack**: the upstream checkpoint is multimodal
  (`Qwen3_5MoeForConditionalGeneration`); mlx-lm's `qwen3_5_moe` loader
  drops the vision tower (`model.visual.*`) by design, so this pack
  ships only the text MoE. Use the upstream repo if you need vision.

## Smoke gate

Before upload this pack passed a deterministic coherence gate: greedy
64-token chat generation loaded through `mlx_lm.load`,
judged for emptiness, repetition loops, multi-script gibberish, and
special-token debris. Verdict: `ok`.

## Usage

```bash
pip install mlx-lm
mlx_lm.generate --model majentik/KAT-Coder-V2.5-Dev-MLX-4bit --prompt "Write a binary search in Python"
```

## Evaluation

| Benchmark | Score |
| --- | --- |
| arc_easy_acc | 0.7000 |
| hellaswag_acc | 0.5400 |

## Available tiers

| Repo | Quant |
| --- | --- |
| [majentik/KAT-Coder-V2.5-Dev-MLX-4bit](https://huggingface.co/majentik/KAT-Coder-V2.5-Dev-MLX-4bit) | 4bit |
| [majentik/KAT-Coder-V2.5-Dev-MLX-6bit](https://huggingface.co/majentik/KAT-Coder-V2.5-Dev-MLX-6bit) | 6bit |
| [majentik/KAT-Coder-V2.5-Dev-MLX-8bit](https://huggingface.co/majentik/KAT-Coder-V2.5-Dev-MLX-8bit) | 8bit |
| [majentik/KAT-Coder-V2.5-Dev-MLX-MXFP4](https://huggingface.co/majentik/KAT-Coder-V2.5-Dev-MLX-MXFP4) | MXFP4 |