---
title: LFM2.5-350M-OptiQ-4bit
canonical_url: "https://www.modelscope.cn/models/mlx-community/LFM2.5-350M-OptiQ-4bit"
md_url: "https://www.modelscope.cn/models/mlx-community/LFM2.5-350M-OptiQ-4bit.md"
repository: mlx-community/LFM2.5-350M-OptiQ-4bit
last_updated: 2026-10-15
license: other
pipeline_tag: text-generation
tasks:
  - text-generation
model_type:
  - lfm2
architectures:
  - Lfm2ForCausalLM
base_model:
  - LiquidAI/LFM2.5-350M
base_model_relation: quantized
parameters: 76.0M
tensor_type:
  - BF16
  - U32
library_name:
  - mlx
  - safetensors
  - pytorch
frameworks:
  - pytorch
downloads: 26
stars: 0
tags:
  - mlx
  - optiq
  - quantized
  - 4bit
  - mixed-precision
  - lfm2
  - liquid-ai
  - apple-silicon
  - on-device
---

# LFM2.5-350M-OptiQ-4bit

> LFM2.5-350M-OptiQ-4bit - mlx-community 在 ModelScope 开源的模型。mlx-community/LFM2.5-350M-OptiQ-4bit

mlx-community/LFM2.5-350M-OptiQ-4bit 是 ModelScope 魔搭社区上的 76.0M 参数text-generation模型，采用 other 许可，基于 LiquidAI/LFM2.5-350M 构建。

- **Repository**: mlx-community/LFM2.5-350M-OptiQ-4bit
- **License**: other
- **Tasks**: text-generation
- **Parameters**: 76.0M
- **Base model**: LiquidAI/LFM2.5-350M
- **Tags**: mlx, optiq, quantized, 4bit, mixed-precision, lfm2, liquid-ai, apple-silicon, on-device
- **Downloads**: 26
- **Stars**: 0
- **Last updated**: 2026-10-15

Source: https://www.modelscope.cn/models/mlx-community/LFM2.5-350M-OptiQ-4bit

---

# mlx-community/LFM2.5-350M-OptiQ-4bit

> **Built with [mlx-optiq](https://mlx-optiq.com)**, the MLX-native toolkit to quantize, fine-tune, and serve LLMs locally on Apple Silicon, no PyTorch and no cloud. [All OptiQ quants](https://mlx-optiq.com/models) · [Docs](https://mlx-optiq.com/docs/)

An OptiQ mixed-precision quant of [LiquidAI/LFM2.5-350M](https://huggingface.co/LiquidAI/LFM2.5-350M), a 350M on-device model from Liquid AI. 269 MB on disk, down from 709 MB at bf16.

LFM2.5 is a hybrid architecture: 16 blocks alternating short convolutions with full attention (only 6 blocks carry a KV cache). OptiQ measures each layer's sensitivity and assigns per-layer bit-widths, so the layers that matter keep 8-bit while the rest go to 4-bit.

## What it is

| Property | Value |
|---|---|
| Base | [LiquidAI/LFM2.5-350M](https://huggingface.co/LiquidAI/LFM2.5-350M) |
| Architecture | `lfm2` — 16 blocks, 10 conv + 6 full-attention |
| Method | OptiQ mixed-precision, sensitivity-driven (bf16 reference) |
| On disk | 269 MB (bf16: 709 MB) |
| Context | 128k |
| KV cache | mixed-precision `kv_config.json` bundled, covering the 6 attention layers |

## Capability Score

Six-metric mean, the standard OptiQ eval.

| Metric | Score |
|---|---:|
| MMLU (5-shot, 969 samples) | 30.3% |
| GSM8K (1000 samples) | 2.5% |
| IFEval (full set, strict) | 69.5% |
| BFCL-V3 simple (200 calls) | 79.5% |
| HumanEval (164 problems, pass@1) | 15.2% |
| HashHop (long-context retrieval) | 0.0% |
| **Capability Score** (mean of 6) | **32.85** |

A 350M model sits near the floor on multi-step maths and long-context retrieval; those scores are genuine, not harness artifacts. Instruction-following and tool-calling are where a model this size is actually useful.

## Run it

```bash
pip install mlx-optiq
```

```python
from mlx_lm import load, generate

model, tok = load("mlx-community/LFM2.5-350M-OptiQ-4bit")
prompt = tok.apply_chat_template(
    [{"role": "user", "content": "Name three colours."}],
    tokenize=False, add_generation_prompt=True,
)
print(generate(model, tok, prompt=prompt, max_tokens=200))
```

Serve it with the bundled mixed-precision KV cache:

```bash
optiq serve --model mlx-community/LFM2.5-350M-OptiQ-4bit --kv-config kv_config.json
```

## Tool calling

LFM2.5 writes Pythonic calls between special tokens:

```
<|tool_call_start|>[get_weather(city="Paris")]<|tool_call_end|>
```

Pass `tools=` to `apply_chat_template` and they are rendered into the system prompt.

## Links

- **Project website:** [mlx-optiq.com](https://mlx-optiq.com/)
- **All OptiQ quants:** [mlx-optiq.com/models](https://mlx-optiq.com/models)
- **Base model:** [LiquidAI/LFM2.5-350M](https://huggingface.co/LiquidAI/LFM2.5-350M)