Liquid AI's LFM2.5-8B-A1B brings 128K context and 38T tokens to edge devices

Liquid AI has launched LFM2.5-8B-A1B, the latest iteration of its edge AI model designed for real-world applications. Building on LFM2-8B-A1B, this version introduces three major upgrades: an expanded 128K context window, training on 38 trillion tokens (up from 12T), and large-scale reinforcement learning. The model also doubles its vocabulary, improving tokenization efficiency for non-Latin languages like Chinese, Arabic, and Hindi. Despite these enhancements, the model remains lightweight enough to run on an entry-level consumer laptop, making it one of the most capable edge models available.

This release signals a shift toward powerful, locally executable AI that doesn't sacrifice advanced features. With 128K context, users can process entire codebases, long documents, or multi-turn conversations without cloud reliance. The reinforcement learning training improves tool-calling capabilities, allowing the model to chain multiple API calls or automate complex workflows autonomously. Liquid AI has made the model freely available on Hugging Face, enabling developers to deploy it in privacy-sensitive or offline environments. For professionals needing AI that runs entirely on their hardware while handling sophisticated tasks, LFM2.5-8B-A1B offers a compelling balance of performance and portability.