Liquid AI has launched LFM2.5-350M, an upgraded model featuring enhanced pre-training with a jump from 10 trillion to 28 trillion tokens, alongside large-scale reinforcement learning. Built on the LFM2 architecture, this model offers fast inference across various hardware, from cloud GPUs to budget-friendly CPUs. It specializes in tasks like data extraction and tool use, making it suitable for large-scale data processing and function calling in edge applications. The model significantly outperforms its predecessor and larger models on key benchmarks. For example, instruction following improved from an IFBench score of 18.20 to 40.69, and data extraction rose from 11.67 to 32.45 on CaseReportBench. Liquid AI has partnered with companies like Distil Labs to fine-tune the model for specific applications, achieving over 95% accuracy in multi-turn interactions for smart home and banking use cases. To support deployment, LFM2.5-350M has day-one compatibility with various inference platforms, including Liquid's LEAP for mobile and ONNX for cross-platform use. Performance benchmarks show it achieves a peak throughput of 40.4K output tokens per second on an NVIDIA H100 GPU. Partnerships with firms like AMD, Qualcomm, and Intel ensure the model is optimized for different hardware environments, enabling effective use on diverse devices, including low-cost smartphones and edge systems.