Liquid AI releases 2.6B-parameter open-source agent model LFM2.5-2.6B for on-device use

AI startup Liquid AI released its open-source agent model LFM2.5-2.6B on August 4. With just 2.6 billion parameters, it can run on devices like smartphones while supporting planning, calling, and multi-step agent workflows.

The model was pre-trained on a 34-trillion-token dataset, and its tokenizer was expanded to better support non-Latin languages. Liquid AI offers both base and post-trained versions on Hugging Face.

In benchmarks against larger models like Gemma-4-E2B-it, Gemma-4-E4B-it, Qwen3.5-4B, and Qwen3.5-9B, LFM2.5-2.6B leads in all instruction-following tasks and nearly all tool-use tests, only trailing Qwen3.5-9B in BFCLv4. In STEM, it tops the AA Omniscience test but lags behind Qwen3.5-9B in math. Coding remains the only area where larger models still have an edge.