Liquid AI Launches Liquid Foundation Models 2.5 230M

Liquid AI has launched the smallest and fastest LFM2.5-230M model to help developers build Agentic workflows. Designed based on the LFM2 architecture, this model can run in a variety of environments from cloud GPUs to low-spec CPUs, delivering fast inference speeds of 213 tok/s on the Galaxy S25 Ultra and 42 tok/s on the Raspberry Pi 5.

This model was pretrained on 19T tokens, and underwent the following 3-stage post-training process to increase developer usability:

• Supervised fine-tuning (SFT) based on distillation using the LFM2.5-350M model
• Direct Preference Optimization (DPO)
• Multi-domain reinforcement learning (RL)

In particular, LFM2.5-230M was used as an on-device control layer for the Unitree G1 humanoid robot, demonstrating the ability to decompose natural language commands into complex tool-calling sequences. According to benchmark results, this model showed performance that competes with or surpasses models more than twice its size on key metrics such as GPQA Diamond and MMLU-Pro.