Liquid AI Open-Sources LFM2.5-2.6B: Compact On-Device Agentic Model With 128K Context and Competitive Benchmarks

Liquid AI has open-sourced LFM2.5-2.6B, a 2.69-billion-parameter agentic model designed to run entirely on-device. The model features a 131,072-token context window, supports tool calling and multi-step reasoning on phones, laptops, and robots, and requires minimal inference cost since processing stays local. Pre-trained on approximately 34 trillion tokens, it achieves instruction-following and tool-use scores competitive with models nearly four times its size, though it trails on coding benchmarks. Both the base checkpoint and a post-trained version are publicly available on Hugging Face under the lfm1.0 license, with native support in llama.cpp, vLLM, SGLang, and LM Studio.