Liquid AI's LFM2-24B-A2B hits 50 tok/s in browser

This is the kind of number that makes browser AI feel less like a stunt and more like a real deployment path.

• –The MoE setup matters: roughly 24B total parameters but only about 2B active per token keeps per-step compute low enough for client-side inference.
• –WebGPU plus Transformers.js/ONNX removes the server hop, which is a real win for privacy, latency, and offline-capable apps.
• –The 8B A1B result above 100 tok/s is the more immediately shippable target for interactive tools, while the 24B run shows the family can scale without falling off a speed cliff.
• –A public Space plus published weights/source lowers the barrier for developers who want to fork the stack and build local assistants or tool-using UIs.
• –Caveat: these numbers come from a high-end Mac, so real-world throughput will vary with browser, quantization, context length, and prompt length.