---
title: LFM2.5-230M-OptiQ-4bit
canonical_url: "https://www.modelscope.cn/models/mlx-community/LFM2.5-230M-OptiQ-4bit"
md_url: "https://www.modelscope.cn/models/mlx-community/LFM2.5-230M-OptiQ-4bit.md"
repository: mlx-community/LFM2.5-230M-OptiQ-4bit
last_updated: 2026-10-15
license: other
pipeline_tag: text-generation
tasks:
  - text-generation
model_type:
  - lfm2
architectures:
  - Lfm2ForCausalLM
base_model:
  - LiquidAI/LFM2.5-230M
base_model_relation: quantized
parameters: 50.7M
tensor_type:
  - BF16
  - U32
library_name:
  - mlx
  - safetensors
  - pytorch
frameworks:
  - pytorch
downloads: 20
stars: 0
tags:
  - mlx
  - optiq
  - quantized
  - 4bit
  - mixed-precision
  - lfm2
  - liquid-ai
  - apple-silicon
  - on-device
---

# LFM2.5-230M-OptiQ-4bit

> LFM2.5-230M-OptiQ-4bit - mlx-community 在 ModelScope 开源的模型。mlx-community/LFM2.5-230M-OptiQ-4bit

mlx-community/LFM2.5-230M-OptiQ-4bit 是 ModelScope 魔搭社区上的 50.7M 参数text-generation模型，采用 other 许可，基于 LiquidAI/LFM2.5-230M 构建。

- **Repository**: mlx-community/LFM2.5-230M-OptiQ-4bit
- **License**: other
- **Tasks**: text-generation
- **Parameters**: 50.7M
- **Base model**: LiquidAI/LFM2.5-230M
- **Tags**: mlx, optiq, quantized, 4bit, mixed-precision, lfm2, liquid-ai, apple-silicon, on-device
- **Downloads**: 20
- **Stars**: 0
- **Last updated**: 2026-10-15

Source: https://www.modelscope.cn/models/mlx-community/LFM2.5-230M-OptiQ-4bit

---

# mlx-community/LFM2.5-230M-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/LFM2.5-230M](https://huggingface.co/LiquidAI/LFM2.5-230M), the smallest model in the LFM2.5 line. 180 MB on disk, down from 459 MB at bf16, small enough to sit in memory on any Apple Silicon Mac alongside everything else you are running.

LFM2.5 is a hybrid architecture: convolutional blocks interleaved with full attention, so only a few blocks carry a KV cache. OptiQ measures each layer's sensitivity and assigns per-layer bit-widths.

## What it is

| Property | Value |
|---|---|
| Base | [LiquidAI/LFM2.5-230M](https://huggingface.co/LiquidAI/LFM2.5-230M) |
| Architecture | `lfm2` — hybrid conv + full attention |
| Method | OptiQ mixed-precision, sensitivity-driven (bf16 reference) |
| On disk | 180 MB (bf16: 459 MB) |
| Context | 128k |
| KV cache | mixed-precision `kv_config.json` bundled |

## Capability Score

Six-metric mean, the standard OptiQ eval.

| Metric | Score |
|---|---:|
| MMLU (5-shot, 969 samples) | 34.2% |
| GSM8K (1000 samples) | 26.0% |
| IFEval (full set, strict) | 60.4% |
| BFCL-V3 simple (200 calls) | 30.0% |
| HumanEval (164 problems, pass@1) | 10.4% |
| HashHop (long-context retrieval) | 0.0% |
| **Capability Score** (mean of 6) | **26.83** |

At 230M parameters the strengths are instruction-following and maths for the size. Long-context multi-hop retrieval sits at the floor for a model this small, and that score is genuine rather than a harness artifact.

## Run it

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

```python
from mlx_lm import load, generate

model, tok = load("mlx-community/LFM2.5-230M-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/LFM2.5-230M-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/LFM2.5-230M](https://huggingface.co/LiquidAI/LFM2.5-230M)