Introducing the on-device-focused open model 'LFM2.5' series, including the small AI model 'LFM2.5-1.2B-JP' specialized for Japanese
AI development company Liquid AI released the LFM2.5 , a small AI model focused on local operation on devices, on January 5, 2026. The LFM2.5 series also includes the Japanese-specific LFM2.5-1.2B-JP, and all models are available for free download.
Introducing LFM2.5: The Next Generation of On-Device AI | Liquid AI
The LFM2.5 series comes in five models: the base model 'LFM2.5-1.2B-Base,' the additionally trained 'LFM2.5-1.2B-Instruct,' the Japanese-specialized 'LFM2.5-1.2B-JP,' the visual language model 'LFM2.5-VL-1.6B,' and the audio language model 'LFM2.5-Audio-1.5B,' and all are released as open models.
LFM2.5-1.2B-Instruct is a lightweight AI model that can run on laptops and smartphones. It can output 239 tokens per second on a PC equipped with an AMD Ryzen AI Max+ 395 processor, and 71 tokens per second on a Samsung Galaxy S25 Ultra.
Below is a table summarizing the various benchmark results for 'LFM2.5-1.2B-Instruct' and the open models 'Llama 3.2 1B Instruct,' 'Gemma 3 1BIT,' 'Granite-4.0-h-1b,' 'Granite-4.0-1b,' and 'Qwen3-1.7B.' LFM2.5-1.2B-Instruct shows the highest performance among models of the same size.
LFM2.5-1.2B-JP is a model specialized for Japanese language processing. It is notable that it only offers a Japanese-specific model, and does not offer specialized models for other languages. In benchmarks measuring performance in Japanese environments, it has recorded significantly higher scores than models of the same size.
The visual language model LFM2.5-VL-1.6B also recorded higher image recognition performance than other models.
The speech language model is designed with quantization in mind, and the lightweight FP4 model achieves performance equivalent to the previous generation FP32 model.
The various models are available at the following links:
LiquidAI/LFM2.5-1.2B-Base · Hugging Face
LiquidAI/LFM2.5-1.2B-Instruct · Hugging Face
https://huggingface.co/LiquidAI/LFM2.5-1.2B-Instruct
LiquidAI/LFM2.5-1.2B-JP · Hugging Face
https://huggingface.co/LiquidAI/LFM2.5-1.2B-JP
LiquidAI/LFM2.5-VL-1.6B · Hugging Face
https://huggingface.co/LiquidAI/LFM2.5-VL-1.6B
LiquidAI/LFM2.5-Audio-1.5B · Hugging Face
https://huggingface.co/LiquidAI/LFM2.5-Audio-1.5B
Liquid AI CEO Ramin Hasani also appeared at AMD's CES keynote on January 6, 2026 (Japan time) and promoted LFM2.5 alongside CEO Lisa Su. The explanation of LFM2.5 during the keynote can be seen from 53 minutes and 30 seconds in the video below.
AMD at CES® 2026 - YouTube
Hasani also revealed plans to release LFM3 in 2026. LFM3 is optimized to run locally on an AI PC equipped with an AMD processor, and can be run as an agent that recognizes voice and images in real time.
Related Posts:
