A compact AI model with Japanese language support that also works on smartphones, the 'LFM2.5-8B-A1B,' has been released.
AI company Liquid AI released its compact language model ' LFM2.5-8B-A1B ' as an open model on May 28, 2026. LFM2.5-8B-A1B can run on smartphones and provide high-quality responses to Japanese questions.
LFM2.5-8B-A1B: An Even Better On-Device Mixture of Experts | Liquid AI
LFM2.5-8B-A1B is a MoE model developed based on ' LFM2-8B-A1B ,' with a total of 8.3 billion parameters and 1.5 billion active parameters. The graph below compares the benchmark results of 'LFM2.5-8B-A1B,' 'Granite-4.0-H-Tiny,' 'gpt-oss-20b,' 'Gemma-4-26B-A4B-IT,' and 'Qwen-3-30B-A3B-Thinking-2507.' Despite being an 8B model, LFM2.5-8B-A1B achieves scores equivalent to models with 20B or more.
When the LFM2.5-8B-A1B is run on a PC equipped with 'AMD Ryzen AI Max+ 395,' it can achieve a high output of 148 tokens per second. Furthermore, even on a Samsung smartphone, 'Galaxy S26 Ultra (SM-S948U1),' a practical output of 28 tokens per second is possible.
Furthermore, the tokenizer has been improved, allowing for efficient tokenization of Japanese, English, Arabic, German, Spanish, French, Hindi, Indonesian, Italian, Korean, Polish, Portuguese, Russian, Thai, Vietnamese, and Chinese. The tokenization efficiency for Japanese has improved by 6.9% compared to the previous tokenizer.
A demo app is available that allows you to experience the performance of the LFM2.5-8B-A1B, so let's actually test its Japanese response performance.
LFM2.5 8B A1B - a Hugging Face Space by LiquidAI
When I typed, 'Should I put a screen protector on my smartphone? Or is it okay not to?', my thoughts started in English.
After the thought process was complete, a response was displayed in Japanese. With other models of a similar size to the LFM2.5-8B-A1B, problems such as 'not being able to output in Japanese at all' or 'the response being incoherent' tend to occur, but the LFM2.5-8B-A1B accurately presented the advantages and disadvantages of applying a screen protector.
LFM2.5-8B-A1B is released as an open model, and the base model, GGUF version, ONNX version, and MLX version are available for download at the following link.
LiquidAI/LFM2.5-8B-A1B · Hugging Face
https://huggingface.co/LiquidAI/LFM2.5-8B-A1B
LiquidAI/LFM2.5-8B-A1B-GGUF · Hugging Face
https://huggingface.co/LiquidAI/LFM2.5-8B-A1B-GGUF
LiquidAI/LFM2.5-8B-A1B-ONNX · Hugging Face
https://huggingface.co/LiquidAI/LFM2.5-8B-A1B-ONNX
LiquidAI/LFM2.5-8B-A1B-MLX-8bit · Hugging Face
https://huggingface.co/LiquidAI/LFM2.5-8B-A1B-MLX-8bit
Related Posts:
in AI, Posted by log1o_hf