---
library_name: transformers
license: other
license_name: lfm1.0
license_link: LICENSE
language:
- en
- ar
- zh
- fr
- de
- ja
- ko
- es
pipeline_tag: text-generation
tags:
- liquid
- lfm2.5
- edge
- gguf
---
# LFM2.5-1.2B-Instruct (W.I.P.)
LFM2.5 is a new family of hybrid models designed for on-device deployment. It builds on the LFM2 architecture with extended pre-training and reinforcement learning.
**Highlights**
* **Best performance** among sub-2B models, particularly in instruction following.
* **2x faster inference** on CPU compared to Qwen3, with optimized prefill and decode speeds.
* **Hybrid architecture** with a combination of convolution and attention blocks.
## LFM2.5 Family
In the LFM2.5 family, we release:
| Model | Description |
|-------|-------------|
| [LFM2.5-1.2B-Base](https://huggingface.co/LiquidAI/LFM2.5-1.2B-Base) | Pre-trained base model |
| [**LFM2.5-1.2B-Instruct**](https://huggingface.co/LiquidAI/LFM2.5-1.2B-Instruct) | General-purpose chat model |
| [LFM2.5-1.2B-JP](https://huggingface.co/LiquidAI/LFM2.5-1.2B-JP) | Japanese-optimized chat model |
| [LFM2.5-VL-1.6B](https://huggingface.co/LiquidAI/LFM2.5-VL-1.6B) | Vision-language model |
| [LFM2.5-1.5B-Audio](https://huggingface.co/LiquidAI/LFM2.5-1.5B-Audio) | Audio-language model |
## Model Details
| Property | Value |
|----------|-------|
| Parameters | 1.17B |
| Context length | 32,768 tokens |
| Architecture | 16 layers (10 conv + 6 attn) |
| Vocabulary | 65,536 |
| Precision | bfloat16 |
| Training budget | 10T tokens |
| License | [LFM Open License v1.0](LICENSE) |
**Supported languages**: English, Arabic, Chinese, French, German, Japanese, Korean, Spanish
**Recommended use cases**: Agentic tasks, data extraction, RAG, creative writing, multi-turn conversations
**Not recommended for**: Knowledge-intensive tasks, programming
## Quick Start
```bash
pip install -U transformers
```
```python
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "LiquidAI/LFM2.5-1.2B-Instruct"
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto", torch_dtype="bfloat16")
tokenizer = AutoTokenizer.from_pretrained(model_id)
messages = [{"role": "user", "content": "What is C. elegans?"}]
input_ids = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
output = model.generate(
input_ids,
do_sample=True,
temperature=0.3,
min_p=0.15,
repetition_penalty=1.05,
max_new_tokens=512,
)
print(tokenizer.decode(output[0], skip_special_tokens=True))
```
**Recommended generation parameters**: `temperature=0.3`, `min_p=0.15`, `repetition_penalty=1.05`
## Chat Template
LFM2.5 uses a ChatML-like format. See the [Chat Template documentation](https://docs.liquid.ai/lfm/key-concepts/chat-template) for details.
```
<|startoftext|><|im_start|>system
You are a helpful assistant trained by Liquid AI.<|im_end|>
<|im_start|>user
What is C. elegans?<|im_end|>
<|im_start|>assistant
```
Use `tokenizer.apply_chat_template()` to automatically format your messages.
## Tool Use
LFM2.5 supports function calling. See the [Tool Use documentation](https://docs.liquid.ai/lfm/key-concepts/tool-use) for the full guide.
```python
tools = [{
"name": "get_weather",
"description": "Get current weather for a location",
"parameters": {
"type": "object",
"properties": {"location": {"type": "string"}},
"required": ["location"]
}
}]
messages = [{"role": "user", "content": "What's the weather in Paris?"}]
input_ids = tokenizer.apply_chat_template(messages, tools=tools, add_generation_prompt=True, return_tensors="pt")
```
## Inference
| Framework | Documentation |
|-----------|--------------|
| Transformers | [docs.liquid.ai/lfm/inference/transformers](https://docs.liquid.ai/lfm/inference/transformers) |
| vLLM | [docs.liquid.ai/lfm/inference/vllm](https://docs.liquid.ai/lfm/inference/vllm) |
| llama.cpp | [LFM2.5-1.2B-Instruct-GGUF](https://huggingface.co/LiquidAI/LFM2.5-1.2B-Instruct-GGUF) |
| Ollama | [docs.liquid.ai/lfm/inference/ollama](https://docs.liquid.ai/lfm/inference/ollama) |
## Fine-tuning
We recommend fine-tuning on your specific use case for best results.
| Method | Link |
|--------|------|
| SFT with Unsloth | [Colab Notebook](https://colab.research.google.com/drive/1HROdGaPFt1tATniBcos11-doVaH7kOI3?usp=sharing) |
| SFT with TRL | [Colab Notebook](https://colab.research.google.com/drive/1j5Hk_SyBb2soUsuhU0eIEA9GwLNRnElF?usp=sharing) |
| DPO with TRL | [Colab Notebook](https://colab.research.google.com/drive/1MQdsPxFHeZweGsNx4RH7Ia8lG8PiGE1t?usp=sharing) |
See the [Fine-tuning documentation](https://docs.liquid.ai/lfm/fine-tuning/trl) for more details.
## Benchmarks
| Model | GPQA | MMLU-Pro | IFEval | IFBench | Multi-IF | AIME25 | BFCLv3 |
|-------|------|----------|--------|---------|----------|--------|--------|
| **LFM2.5-1.2B-Instruct (BF16)** | 35.81 | 44.76 | 85.76 | 47.33 | 61.22 | 12.33 | 49.04 |
| Qwen3-1.7B | 34.85 | 42.91 | 73.68 | 21.33 | 56.48 | 9.33 | 46.30 |
| Granite 4.0-1B | 24.24 | 33.53 | 79.61 | 21.00 | 43.65 | 3.33 | 52.43 |
| Llama 3.2 1B Instruct | 16.57 | 20.80 | 52.37 | 15.93 | 30.16 | 0.33 | 21.44 |
| Gemma 3 1B IT | 24.24 | 14.04 | 63.25 | 20.47 | 44.31 | 1.00 | 16.64 |
## Resources
- [Documentation](https://docs.liquid.ai/lfm)
- [Cookbook](https://github.com/Liquid4All/cookbook)
- [Discord](https://discord.gg/liquid-ai)
- [Blog Post](https://www.liquid.ai/blog/)
## Contact
For enterprise solutions and edge deployment, contact [sales@liquid.ai](mailto:sales@liquid.ai).
## Citation
```bibtex
@article{liquidai2025lfm2,
title={LFM2 Technical Report},
author={Liquid AI},
journal={arXiv preprint arXiv:2511.23404},
year={2025}
}
``` 