Liquid AI's new LFM2.5 8B-A1B MoE model delivers 253 tokens/s on M5 Max, runs under 6GB memory on mobile, and achieves 18,500 output tokens/s on H100—all while outperforming similarly-sized dense models on key benchmarks.
A smaller release drew outsized attention on LocalLLaMA because LFM2.5-350M is not trying to be a general-purpose chatbot. Liquid AI is pitching it as a compact model for tool use, structured outputs, and data-heavy edge workflows.
A LocalLLaMA post claiming that Liquid AI’s LFM2-24B-A2B can run at roughly 50 tokens per second in a browser on an M4 Max reached 79 points and 11 comments. Community interest centered on sparse MoE architecture, ONNX packaging, and whether WebGPU can make the browser a credible local AI deployment target.