A 149 MB LLM, running in your tab.

LFM2.5-230M in the BRIK format, run by hand-written WGSL kernels on your GPU. No inference server, no API key: the text never leaves the machine.

149 MBstreamed by HTTP ranges, then offline
0 serverseverything runs on the visitor's device
~158 tok/sdecoding, measured on an Apple Silicon laptop