LFM2.5-1.2B-JP-202606

<br> <div align="center"> <img src="https://cdn-uploads.huggingface.co/production/uploads/61b8e2ba285851687028d395/2b08LKpev0DNEk6DlnWkY.png" alt="Liquid AI" style="width: 100%; max-width: 100%; height: auto; display: inline-block; margin-bottom: 0.5em; margin-top: 0.5em;" /> <div style="display: flex; justify-content: center; gap: 0.5em;"> <a href="https://playground.liquid.ai/"><strong>Try LFM</strong></a> • <a href="https://docs.liquid.ai/lfm/getting-started/welcome"><strong>Docs</strong></a> • <a href="https://leap.liquid.ai/"><strong>LEAP</strong></a> • <a href="https://discord.com/invite/liquid-ai"><strong>Discord</strong></a> </div> </div> <br>

🇯🇵 LFM2.5-1.2B-JP-202606

<div align="center"> <img src="https://cdn-uploads.huggingface.co/production/uploads/64618cf05dba83471db2be9b/nhW5KrNVrPIe-3zy-RLLt.png" alt="Liquid AI" width="70%" /> </div>

LFM2.5-1.2B-JP-202606 is our latest general purpose Japanese chat model, delivering significant improvements in knowledge, instruction following, math, code, and tool-use over both the models of comparable size and LFM2.5-1.2B-JP. It sets a new benchmark for state-of-the-art performance in Japanese language understanding. Ideal for developers building Japanese-language applications where cultural and linguistic nuance matter.

LFM2.5-1.2B-JP-202606 は、当社の最新の汎用日本語チャットモデルです。知識、指示追従、数学、コード、ツール使用の各領域において、同規模の他モデルおよび LFM2.5-1.2B-JP の双方を大幅に上回る改善を実現しています。日本語全般における最高水準のベンチマーク性能を発揮します。 文化的・言語的なニュアンスが重要となる日本語アプリケーションを構築する開発者に最適です。

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

📊 Performance

<div align="center"> <img src="https://cdn-uploads.huggingface.co/production/uploads/64618cf05dba83471db2be9b/7gqajAlXAh52nAz85JoQt.png" alt="Liquid AI" width="90%" /> <div style="display: flex; justify-content: center; gap: 0.5em;"> </div> </div>

We compared LFM2.5-1.2B-JP-202606 with relevant sub-2B models on a diverse suite of benchmarks.

<table> <thead> <tr> <th rowspan="2">Model</th> <th rowspan="2">Size</th> <th colspan="5">Knowledge</th> <th colspan="3">Instruction Following</th> <th colspan="3">Math</th> <th>Code</th> <th>Tool Use</th> <th rowspan="2">Domain Avg</th> </tr> <tr> <th>JMMLU‑ProX</th> <th>JMMLU</th> <th>JCulture</th> <th>JGPQA</th> <th>Avg</th> <th>J‑MIFEval</th> <th>JFBench¹</th> <th>Avg</th> <th>J‑GSM8K</th> <th>J‑MATH500</th> <th>Avg</th> <th>JHumanEval+</th> <th>J‑BFCLv3²</th> </tr> </thead> <tbody> <tr> <td><strong>LFM2.5‑1.2B‑JP‑202606</strong></td> <td>1.2B</td> <td>36.23</td><td>54.19</td><td>35.77</td><td>28.69</td><td>38.72</td> <td>79.08</td><td>54.77</td><td>66.93</td> <td>62.20</td><td>62.80</td><td>62.50</td> <td>49.39</td> <td>48.00</td> <td><strong>53.11</strong></td> </tr> <tr> <td>LFM2.5‑1.2B‑Instruct</td> <td>1.2B</td> <td>31.42</td><td>47.61</td><td>28.42</td><td>31.72</td><td>34.79</td> <td>40.44</td><td>36.67</td><td>38.56</td> <td>50.20</td><td>50.00</td><td>50.10</td> <td>28.66</td> <td>46.29</td> <td>39.68</td> </tr> <tr> <td>Qwen3‑1.7B (Instruct)</td> <td>1.7B</td> <td>30.78</td><td>47.67</td><td>33.33</td><td>26.26</td><td>34.51</td> <td>40.29</td><td>36.61</td><td>38.45</td> <td>46.00</td><td>56.40</td><td>51.20</td> <td>47.56</td> <td>52.45</td> <td>44.83</td> </tr> <tr> <td>Granite‑4.0‑1B</td> <td>1.5B</td> <td>15.32</td><td>33.93</td><td>34.38</td><td>24.44</td><td>27.02</td> <td>27.56</td><td>31.26</td><td>29.41</td> <td>42.80</td><td>25.40</td><td>34.10</td> <td>51.22</td> <td>50.57</td> <td>38.46</td> </tr> <tr> <td>Llama‑3.2‑1B‑Instruct</td> <td>1.2B</td> <td>15.91</td><td>33.97</td><td>22.52</td><td>32.32</td><td>26.18</td> <td>24.10</td><td>21.78</td><td>22.94</td> <td>25.20</td><td>11.40</td><td>18.30</td> <td>17.68</td> <td>21.06</td> <td>21.23</td> </tr> <tr> <td>Gemma‑3‑1B‑it</td> <td>1.0B</td> <td>14.12</td><td>34.45</td><td>23.42</td><td>24.24</td><td>24.06</td> <td>26.31</td><td>31.15</td><td>28.73</td> <td>33.60</td><td>15.60</td><td>24.60</td> <td>25.00</td> <td>17.26</td> <td>23.93</td> </tr> <tr> <td>sarashina2.2‑1b‑instruct‑v0.1</td> <td>1.4B</td> <td>18.3</td><td>40.24</td><td>25.53</td><td>26.26</td><td>27.58</td> <td>21.9</td><td>27.41</td><td>24.66</td> <td>44.4</td><td>24.8</td><td>34.60</td> <td>21.95</td> <td>13.86</td> <td>24.53</td> </tr> <tr> <td>TinySwallow‑1.5B‑Instruct</td> <td>1.5B</td> <td>21.51</td><td>47.98</td><td>31.17</td><td>29.29</td><td>32.49</td> <td>36.55</td><td>34.25</td><td>35.40</td> <td>47.2</td><td>22.4</td><td>34.80</td> <td>26.83</td> <td>11.7</td> <td>28.24</td> </tr> <tr> <td>llm‑jp‑3.1‑1.8b‑instruct4</td> <td>1.9B</td> <td>17.44</td><td>43.05</td><td>27.42</td><td>17.68</td><td>26.40</td> <td>33.77</td><td>30.92</td><td>32.35</td> <td>52.8</td><td>17.0</td><td>34.90</td> <td>35.37</td> <td>11.76</td> <td>28.16</td> </tr> <tr> <td>RakutenAI‑2.0‑mini‑instruct</td> <td>1.5B</td> <td>11.46</td><td>31.84</td><td>29.67</td><td>22.22</td><td>23.80</td> <td>28.06</td><td>24.66</td><td>26.36</td> <td>24.8</td><td>11.4</td><td>18.10</td> <td>28.6</td> <td>11.85</td> <td>21.74</td> </tr> </tbody> </table>

¹ JFBench is evaluated using single-instruction prompts. <br> ² quickTestingOSSHandler is used for models that do not support function calling (sarashina2.2‑1b‑instruct‑v0.1, TinySwallow‑1.5B‑Instruct, llm‑jp‑3.1‑1.8b‑instruct4, and RakutenAI‑2.0‑mini‑instruct).

🗒️ Model Details

LFM2.5-1.2B-JP-202606 is a general-purpose text-only model with the following features:

• Number of parameters: 1.17B
• Number of layers: 16 (10 double-gated LIV convolution blocks + 6 GQA blocks)
• Training budget: 31.5T tokens
• Context length: 32,768 tokens
• Vocabulary size: 65,536
• Knowledge cutoff: Mid-2024
• Languages: English, Japanese
• Generation parameters:
  • temperature: 0.1
  • top_k: 50
  • repetition_penalty: 1.05

We recommend using it for agentic workflows, tool use, structured outputs, bilingual English–Japanese assistants, and on-device personal-assistant applications. It is not recommended for knowledge-intensive tasks. It performs best when given clear, explicit instructions that define the task, expected behavior, and output format.

エージェント型ワークフロー、ツール使用、構造化出力、日英バイリンガルアシスタント、オンデバイスのパーソナルアシスタントでの利用を推奨します。一方で、詳細な知識を要するのタスクには推奨されません。タスク内容、期待される動作、出力形式を明確かつ具体的に指示することで、最も高い性能を発揮します。

Chat Template

LFM2.5 uses a ChatML-like format. See the Chat Template documentation for details. Example:

<|startoftext|><|im_start|>system
You are a helpful assistant trained by Liquid AI.<|im_end|>
<|im_start|>user
日本の首都は？<|im_end|>
<|im_start|>assistant

You can use tokenizer.apply_chat_template() to format your messages automatically.

Tool Use

LFM2.5 supports function calling as follows:

1. Function definition: We recommend providing the list of tools as a JSON object in the system prompt. You can also use the tokenizer.apply_chat_template() function with tools.
2. Function call: By default, LFM2.5 writes Pythonic function calls (a Python list between <|tool_call_start|> and <|tool_call_end|> special tokens), as the assistant answer. You can override this behavior by asking the model to output JSON function calls in the system prompt.
3. Function execution: The function call is executed, and the result is returned as a "tool" role.
4. Final answer: LFM2 interprets the outcome of the function call to address the original user prompt in plain text.

See the Tool Use documentation for the full guide. Example:

<|startoftext|><|im_start|>system
List of tools: [{"name": "get_candidate_status", "description": "採用プロセスにおける候補者の現在のステータスを取得します", "parameters": {"type": "object", "properties": {"candidate_id": {"type": "string", "description": "候補者の一意の識別子"}}, "required": ["candidate_id"]}}]<|im_end|>
<|im_start|>user
候補者ID 12345 の現在のステータスは何ですか？<|im_end|>
<|im_start|>assistant
<|tool_call_start|>[get_candidate_status(candidate_id="12345")]<|tool_call_end|>候補者ID 12345 の現在のステータスを確認しています。<|im_end|>
<|im_start|>tool
[{"candidate_id": "12345", "status": "Interview Scheduled", "position": "Clinical Research Associate", "date": "2023-11-20"}]<|im_end|>
<|im_start|>assistant
ID 12345 の候補者は現在、Clinical Research Associate のポジションで「面接予定」の段階にあり、面接日は 2023年11月20日に設定されています。<|im_end|>

🏃 Inference

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

Here's a quick start example with Transformers:

from transformers import AutoModelForCausalLM, AutoTokenizer, TextStreamer
model_id = "LFM2.5-1.2B-JP-202606"
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 = "日本の首都は？"
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=0.1,
    top_k=50,
    repetition_penalty=1.05,
    max_new_tokens=512,
    streamer=streamer,
)

🔧 Fine-Tuning

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

📬 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}
}