---
title: LFM2.5-24B-A2B-Preview-mxfp8
canonical_url: "https://www.modelscope.cn/models/mlx-community/LFM2.5-24B-A2B-Preview-mxfp8"
md_url: "https://www.modelscope.cn/models/mlx-community/LFM2.5-24B-A2B-Preview-mxfp8.md"
repository: mlx-community/LFM2.5-24B-A2B-Preview-mxfp8
last_updated: 2026-10-14
license: other
pipeline_tag: text-generation
tasks:
  - text-generation
model_type:
  - lfm2_moe
architectures:
  - Lfm2MoeForCausalLM
base_model:
  - lq-preview/LFM2.5-24B-A2B-Preview
base_model_relation: quantized
library_name:
  - mlx
  - safetensors
  - pytorch
frameworks:
  - pytorch
language:
  - en
  - ar
  - zh
  - fr
  - de
  - ja
  - ko
  - es
  - pt
  - it
downloads: 42
stars: 0
tags:
  - mlx
  - mlx-vlm
  - lfm2.5
  - moe
---

# LFM2.5-24B-A2B-Preview-mxfp8

> LFM2.5-24B-A2B-Preview-mxfp8 - mlx-community 在 ModelScope 开源的模型。LFM2.5-24B-A2B-Preview — mxfp8

mlx-community/LFM2.5-24B-A2B-Preview-mxfp8 是 ModelScope 魔搭社区上的text-generation模型，采用 other 许可，基于 lq-preview/LFM2.5-24B-A2B-Preview 构建。

- **Repository**: mlx-community/LFM2.5-24B-A2B-Preview-mxfp8
- **License**: other
- **Tasks**: text-generation
- **Base model**: lq-preview/LFM2.5-24B-A2B-Preview
- **Tags**: mlx, mlx-vlm, lfm2.5, moe
- **Downloads**: 42
- **Stars**: 0
- **Last updated**: 2026-10-14

Source: https://www.modelscope.cn/models/mlx-community/LFM2.5-24B-A2B-Preview-mxfp8

---

# LFM2.5-24B-A2B-Preview — mxfp8

MLX conversion of [lq-preview/LFM2.5-24B-A2B-Preview](https://huggingface.co/lq-preview/LFM2.5-24B-A2B-Preview), pinned to revision `f772950ce6f17d35af5469d00334b61a9c2fc034`. Converted with the MLX-VLM CLI on an Apple M3 Ultra. This is a text generation model.

MXFP8 weight-only quantization, 8 bits, group size 32. Router projections use affine 8-bit quantization (group size 64). Unquantized floating-point weights remain BF16; expert routing biases remain FP32.

The source configuration specifies a context limit of **131,072 tokens**. Weight files total **24.72 GB** (decimal); runtime memory also includes cache and temporary allocations.

## Usage

This export was tested with the LFM2.5 port in [MLX-VLM PR #2505](https://github.com/Blaizzy/mlx-vlm/pull/2505), commit `a6c7313bc053e5887616fa850808d21576eddf98`, MLX 0.32.3 and Transformers 5.17.0.

```bash
pip install 'git+https://github.com/Blaizzy/mlx-vlm.git@a6c7313bc053e5887616fa850808d21576eddf98'
python -m mlx_vlm generate \
  --model nativ-community/LFM2.5-24B-A2B-Preview-mxfp8 \
  --prompt "What is the capital of France? Answer briefly." \
  --max-tokens 512 \
  --temperature 0.1
```

Identical model files are published in both organizations:

- [nativ-community/LFM2.5-24B-A2B-Preview-mxfp8](https://huggingface.co/nativ-community/LFM2.5-24B-A2B-Preview-mxfp8)
- [mlx-community/LFM2.5-24B-A2B-Preview-mxfp8](https://huggingface.co/mlx-community/LFM2.5-24B-A2B-Preview-mxfp8)

Use the included chat template. The model generates reasoning before its final answer, so allow enough output tokens.

## Conversion

```bash
python -m mlx_vlm convert --hf-path lq-preview/LFM2.5-24B-A2B-Preview --revision f772950ce6f17d35af5469d00334b61a9c2fc034 --mlx-path ./LFM2.5-24B-A2B-Preview-mxfp8 --dtype bfloat16 -q --q-mode mxfp8 --q-bits 8 --q-group-size 32
```

The export was reloaded with strict weight matching. Prefill and cached decode produced finite logits, and two short greedy generation checks returned Paris and 391 correctly and reached EOS. These checks are smoke tests, not a comprehensive quality evaluation or a long-context benchmark. See `conversion_info.json` for settings and versions.

## License and attribution

The original model is by Liquid AI, Inc. The original [LFM Open License v1.0](LICENSE) is included and applies to these weights. Conversion and quantization modifications are described in [NOTICE](NOTICE).