Liquid AI has released LFM2.5-2.6B, a compact model designed to run capable agents on local hardware instead of depending on cloud inference. The Hugging Face post says the model supports tool calling and multi-step workflows while remaining small enough for laptops, phones, and other everyday devices.
The model was pre-trained on about 34 trillion tokens, then mid-trained to extend context length to 128,000 tokens. Liquid AI says post-training turned the base model into an agent through a pipeline that separates model optimization, inference, and environment execution.
For developers, the main claim is practical deployment. A smaller local model can keep user data on the device, reduce inference bills, and support applications that need responsive tool use without routing every step through a hosted frontier model.
Liquid AI reports that LFM2.5-2.6B competes with or beats several larger models on instruction-following and tool-use benchmarks, while staying close on agentic tasks. Those are vendor-reported results, so real-world testing still matters, especially for long workflows and failure recovery. The release nevertheless reflects a broader push toward smaller agent models that can be embedded directly into products.