---
language:
- en
license: mit
tags:
- Explorer SubAgent
- Repository Exploration
- mlx
library_name: mlx
base_model: microsoft/FastContext-1.0-4B-SFT
pipeline_tag: text-generation
---

# FastContext-1.0-4B-SFT-mlx-4bit-g32

4-bit MLX quantization of [microsoft/FastContext-1.0-4B-SFT](https://huggingface.co/microsoft/FastContext-1.0-4B-SFT) with group_size=32 for Apple Silicon.

## Quantization details

- **Method:** Affine 4-bit
- **Group size:** 32 (finer than the default 64)
- **Effective bits per weight:** 5.0
- **Model size:** 2.4 GB (vs 7.5 GB bf16)

## Benchmark results

Tested on 10 SWE-bench Multilingual instances against other quantization variants:

| Model | Bits/Wt | Size | File F1 | Line F1 |
|-------|---------|------|---------|---------|
| affine 8-bit g64 | 8.5 | 4.0G | 0.507 | 0.140 |
| **affine 4-bit g32 (this model)** | **5.0** | **2.4G** | **0.300** | **0.090** |
| affine 3-bit g64 | 3.5 | 1.7G | 0.100 | 0.000 |
| affine 4-bit g64 | 4.5 | 2.1G | 0.050 | 0.005 |
| mattrobenolt 4-bit g64 | 4.5 | 2.1G | 0.025 | 0.008 |

The finer group_size=32 delivers **12x better File F1** than standard 4-bit g64 quantization with only 300MB additional size.

## Usage

```python
from mlx_lm import load, generate

model, tokenizer = load("rubybear-lgtm/FastContext-1.0-4B-SFT-mlx-4bit-g32")
```

Or with [fastcontext-mcp](https://github.com/rubybear-lgtm/fastcontext) for Claude Code integration.