Chat LFM2.5 350M

Running locally with jax-js + WebGPU

Options

Model

Switching models keeps each checkpoint cached in your browser.

Backend

WebGPU uses fp16 weights. Wasm casts weights to fp32 on load.

Max new tokens Top-k Temperature: 0.10 Top-p: 1.00 Repetition penalty: 1.05

KV cache is allocated dynamically for the current chat.

Talk to an LLM

The first message downloads and caches a 676 MB fp16 checkpoint. Everything runs locally in your browser.