A 2.6-Billion-Parameter Local Agent With LFM2.5

LFM2.5-2.6B runs agents locally on phones or computers, with a 128K context window, tool calling, and open weights.

LFM2.5-2.6B is an open-weights model designed to run agents directly on a phone, computer, or robot. It can plan a task, call tools, and work through multiple steps without sending data to a cloud service.

The model has 2.6 billion parameters and was pretrained on approximately 34 trillion tokens. It supports a 128K context window and a vocabulary of the same size, intended in particular to improve coverage of non-Latin writing systems.

Its post-training combines supervised fine-tuning, the specialization of several expert models, distillation, and reinforcement learning in agentic environments. The final stages were conducted with tools such as Pi, Hermes Agent, and OpenClaw, inside sandboxes designed to verify outcomes and enforce safety requirements.

According to Liquid AI, LFM2.5-2.6B scores 77.83 on ToolSandbox, compared with 76.44 for Qwen3.5-9B. It also reaches 80.07 on Multi-IF and 85.49 on IFStruct. The company acknowledges, however, that larger models retain an advantage on coding tasks and the most complex agentic workflows.

The model supports llama.cpp, MLX, vLLM, SGLang, and ONNX, with integrations for Apple, AMD, Qualcomm, NVIDIA, and Intel chips. In its own tests, Liquid AI measured 220 tokens per second on an M5 Max, 113 on a Ryzen AI Max+ 395, and approximately 30 on a phone, with a memory footprint below 2.5 GB.

LFM2.5-2.6B and its Base version are available on Hugging Face under the LFM2 open-weight license. The model can also be served through an OpenAI-compatible endpoint and connected to different agentic environments.