On August 5, the AI startup Liquid AI officially announced the release of LFM2.5-2.6B, a terminal-side model designed for agent workflows. The model has 2.6 billion parameters and can run entirely locally on terminal devices such as smartphones without relying on cloud API, providing developers with a low-cost, low-latency, and privacy-protected edge-side agent solution.
In terms of technical architecture and training approach, LFM2.5-2.6B is pre-trained on about 34 trillion token, with a vocabulary expanded to 128K. Its post-training process integrates four stages: supervised fine-tuning, teacher specialization, multi-domain strategy distillation, and agent reinforcement learning, giving the model excellent planning, tool calling, and complex multi-step task handling capabilities. Benchmark tests show that the model's performance in following instructions and using tools is comparable to large models nearly four times its size.
Additionally, the model performs well in hardware compatibility and inference performance, achieving a reasoning speed of 220 tokens/second on the M5Max chip, and maintaining a stable response speed of 30 tokens/second on mobile devices. Currently, LFM2.5-2.6B and its base version are fully open-sourced on Hugging Face, and natively support mainstream inference frameworks such as llama.cpp, MLX, and vLLM, marking the practical stage of high-efficiency edge-side agents.
