Liquid AI releases a 230M model optimized for phones, Raspberry Pi, and robots (www.liquid.ai)

Liquid AI has launched its latest model, LFM2.5-230M, which is optimized for mobile devices, Raspberry Pi, and robotics. This lightweight model, designed for swift inference across various platforms, achieves notable decode speeds of 213 tokens per second on the Galaxy S25 Ultra and 42 tokens per second on a Raspberry Pi 5. Built on the innovative LFM2 architecture, it is highly adaptable for developers, being pre-trained with a 19 trillion token dataset and incorporating advanced techniques like supervised fine-tuning and multi-domain reinforcement learning. Early tests reveal its capabilities in tool use and data extraction, making it suitable for on-device applications. The significance of LFM2.5-230M lies in its potential to democratize AI applications, enabling developers to deploy effective AI solutions in resource-constrained environments. By demonstrating that a 230 million parameter model can deliver competitive performance against much larger models, particularly in areas such as instruction following and data extraction, Liquid AI paves the way for more accessible AI technologies. The model supports various frameworks for optimization, ensuring low-latency and efficient processing across diverse hardware, and is positioned to be a powerful tool for agile development, particularly in wearable technology and robotics, while keeping its memory footprint minimal.