Try LFM • Docs • LEAP • Discord

LFM2.5-1.2B-JP

LFM2.5-1.2B-JP is a chat model specifically optimized for Japanese. While LFM2 already supported Japanese as one of eight languages, LFM2.5-JP pushes state-of-the-art on Japanese knowledge and instruction-following at its scale. This model is ideal for developers building Japanese-language applications where cultural and linguistic nuance matter.

Find more information about LFM2.5 in our blog post.

🏃 Inference

LFM2.5 is supported by many inference frameworks. See the Inference documentation for the full list.

Name	Description	Docs	Notebook
Transformers	Simple inference with direct access to model internals.	Link
vLLM	High-throughput production deployments with GPU.	Link
llama.cpp	Cross-platform inference with CPU offloading.	Link

Here's a quick start example with transformers:

from transformers import AutoModelForCausalLM, AutoTokenizer, TextStreamer

model_id = "LiquidAI/LFM2.5-1.2B-JP"
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto",
    dtype="bfloat16",
#   attn_implementation="flash_attention_2" <- uncomment on compatible GPU
)
tokenizer = AutoTokenizer.from_pretrained(model_id)
streamer = TextStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)

prompt = "What is C. elegans?"

input_ids = tokenizer.apply_chat_template(
    [{"role": "user", "content": prompt}],
    add_generation_prompt=True,
    return_tensors="pt",
    tokenize=True,
).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,
    streamer=streamer,
)

• Recommended generation parameters:
  • temperature: 0.3
  • min_p: 0.15
  • repetition_penalty: 1.05

🔧 Fine-Tuning

We recommend fine-tuning LFM2.5 for your specific use case to achieve the best results.

Name	Description	Docs	Notebook
SFT (Unsloth)	Supervised Fine-Tuning with LoRA using Unsloth.	Link
SFT (TRL)	Supervised Fine-Tuning with LoRA using TRL.	Link
DPO (TRL)	Direct Preference Optimization with LoRA using TRL.	Link

📊 Performance

Model	JMMLU	M-IFEval (ja)	GSM8K (ja)
LFM2.5-1.2B-JP	50.7	58.1	56.0
LFM2.5-1.2B-Instruct	47.7	41.8	46.8
Qwen3-1.7B (Instruct mode)	47.7	40.3	46.0
Llama 3.2 1B Instruct	34.0	24.1	25.2
TinySwallow-1.5B-Instruct	48.0	36.5	47.2
Gemma-2-Llama-Swallow-2b-it-v0.1	48.1	33.4	34.4
Gemma-3-1b-it	34.5	26.3	33.6
Granite-4.0-h-1b	42.2	39.3	42.8
Sarashina2.2-1b-instruct-v0.1	40.2	21.9	44.4

Evaluation Notes

• All results are zero-shot evaluations using greedy decoding.
• M-IFEval (ja) scores correspond to the loose evaluation setting.
• JMMLU was evaluated using a prompt format in a similar style to the ArtificialAnalysis methodology (with corresponding parsing logic). The Japanese prompt template used is shown below:

PROMPT_TEMPLATE = """与えられた選択問題に答えてください。回答の最後の行に「答え：{valid_options}」のように出力してください（例：「答え：X」）。

{question}

{options}"""

📬 Contact

• Got questions or want to connect? Join our Discord community
• If you are interested in custom solutions with edge deployment, please contact our sales team.

Citation

@article{liquidai2025lfm2,
  title={LFM2 Technical Report},
  author={Liquid AI},
  journal={arXiv preprint arXiv:2511.23404},
  year={2025}
}