Introducing: a coding agent (Pi) running entirely in your browser using a 2B model on WebGPU 🤯 MiniCPM5-2B + Pi, powered by Transformers.js + WebGPU + 4-bit ONNX weights. All previous attempt to create this failed but MiniCPM5 seems to make it usable. Available now on Hugging Face 👇

Try it here:

is this Pi?? 😂

2B running a coding agent locally is the part that gets me what's it actually usable for, autocomplete-tier stuff or can it hold a real task

wow great work! I have been trying to run MiniCPM5-2B in a browser. Tried Transformers.js/ONNX route but WebLLM/MLC works faster in my experience. Here is my Hugging Face space that uses my own MLC quants.

cool

the browser is becoming the smallest useful agent sandbox

The embedding split is the part that makes this work: 510 MiB of embeddings cut losslessly into four 127.5 MiB tensors so it clears WebGPU's 128 MiB storage binding cap. First download is 1.84 GB into OPFS, and cached reloads pull zero weights.

Does the cat have a name @victormustar

Whats the cat mascot called?

pair Pi with Sogni Creative Agent and you have code plus creative production in one very capable setup 😎

Lots of clustered wallets fuck this

wtf lol

Running a usable 2B model entirely in-browser via WebGPU is a genuinely impressive technical feat, huge congrats on the release.

why not use wllama ?

居然能直接在浏览器跑端侧编码智能体，MiniCP5这个小模型适配得真不错，不用部署本地就能体验，挺方便。

How much of the coding workflow survives once the browser tab loses focus or the WebGPU device gets reclaimed?

oh shit you did it!

running an agent client-side on webgpu is insane progress. testing this now

kinda dope

Awesome! Will give it a try

My god you beat me to it Absurdly marvelous thank you for that

2B on WebGPU that actually finishes a loop is the part most people skip. Cloud still wins the hard tasks. This wins the ones you can keep in a tab.

@maceip NOT WRONG! JUST EARLY!

the interesting part isnt that it codes, its that nothing leaves the machine. a 2b wont out-write a frontier model but it can read things youd never paste into an api. whats the latency per token like on a normal laptop gpu?

wait is it actually coding in the browser or just chatting

crazy

This is so cool! Would you be so kind as to open source the web app? The pi TUI + monaco workspace split view is really nice

Any benchmarks?

yes

transformers.js, webgpu and 4-bit onnx all shipped before and died, so the model was the bottleneck. webllm with the same model should get the same jump. the browser part was never the story.

Browser-local coding agent that actually sticks is rare. How do you cap tool loops + WebGPU memory when the session runs long?