Liquid AI has launched LFM2.5-8B-A1B, an advanced edge model designed for fast, reliable tool calling on consumer hardware, on May 28, 2026. This model features an expanded 128K context window and has been pretrained on 38 trillion tokens, more than tripling the data used in its predecessor, LFM2-8B-A1B, released in October 2025, according to liquid.ai.
The development of LFM2.5-8B-A1B involved scaling up pretraining from 12 trillion to 38 trillion tokens and doubling the vocabulary to enhance tokenization efficiency for non-Latin languages. The model supports chaining tool calls and performing complex tasks while running efficiently on entry-level laptops. Both the base and post-trained versions are available on Hugging Face and Liquid AI’s Playground, with documentation provided for local deployment and fine-tuning.
This release is significant as it advances on-device AI capabilities by combining a large context window with efficient mixture-of-experts architecture, enabling sophisticated AI functions without reliance on cloud infrastructure. The model’s improved tokenization for diverse languages broadens its usability globally. Such developments reflect growing demand for privacy-conscious, low-latency AI applications that can operate independently of internet connectivity.
Looking ahead, users can access LFM2.5-8B-A1B immediately and experiment with its capabilities through Liquid AI’s platforms. The company’s documentation supports developers in integrating and customizing the model locally. This release may set a precedent for future edge AI models emphasizing scalability, multilingual support, and efficient resource use on consumer devices.