Liquid AI releases LFM2.5 on-device models
Liquid AI has released LFM2.5, a family of open-weight models of about 1 billion parameters aimed at on-device agents, with self-reported benchmarks that beat Llama 3.2 1B, Gemma 3 1B and Qwen3-1.7B.
Liquid AI announced the LFM2.5-1.2B family, which builds on its LFM2 architecture designed for devices. The company increased pretraining from 10T to 28T tokens and added more reinforcement learning in post-training.
The release includes five models:
- LFM2.5-1.2B-Base, the pretrained checkpoint for heavy fine-tuning
- LFM2.5-1.2B-Instruct, trained with SFT, preference alignment and multi-stage RL
- LFM2.5-1.2B-JP, tuned for Japanese
- LFM2.5-VL-1.6B, a vision-language model
- LFM2.5-Audio-1.5B, a native speech-and-text model
All of them are open-weight and available on Hugging Face and Liquid's LEAP platform. AMD and Nexa AI are launch partners and are providing NPU-optimized performance.
On Liquid's own benchmarks, the Instruct model scores 38.89 on GPQA, 44.35 on MMLU-Pro, 86.23 on IFEval and 14.00 on AIME25. Qwen3-1.7B in instruct mode scores 34.85, 42.91, 73.68 and 9.33 on the same tests. On the BFCLv3 tool-use benchmark (49.12), the Instruct model trails IBM's Granite-4.0-1b (52.43) and Granite-4.0-h-1b (50.69). The Japanese model scores 50.7 on JMMLU, ahead of every comparison model Liquid listed.
The vision-language model improves on its predecessor in multi-image and multilingual tasks. It does slightly worse on RealWorldQA, and InternVL3.5-1B beats it on MMMU and on the text benchmarks GPQA and MMLU-Pro.
Liquid says the audio model is 8x faster than the previous version. It uses a new detokenizer and ships GGUFs that work with llama.cpp for CPU inference. The announcement gives no specific latency or throughput figures, and no outside group has verified the results.