MYGPU runs large language models entirely on your own hardware, inside the browser. No servers. No accounts. Zero data egress.
Load a model once and it lives in your browser. Disconnect from the internet and keep working — there was never a server to lose.
WebGPU runs the model on your graphics card. Every token is computed locally.
After the first load, the model runs with your network completely disconnected.
From a 360M model for speed to a 3B model for depth. Switch without reloading.
Tokens per second, latency, context size and egress — measured, not promised.
Open model weights, 4-bit quantized, downloaded once and cached in your browser.
q4f16_1Compiled WGSL compute shaders dispatch matrix operations straight to your GPU.
MLC · WebGPUThe KV-cache and activations live in your VRAM. We don't operate any servers.
local VRAMTokens stream into your tab. Close it and the session is gone. Nothing is logged.
egress: 0 BYou hold your keys on your own device. Your AI should follow the same rule.
If this website went offline tomorrow, the model already on your machine would keep running.
$MYGPU is the community behind private, self-hosted AI. Not your GPU, not your AI.
No sign-up. Open the terminal, load a model, and start working privately.