What runs
Qwen2.5 0.5B, 4-bit, through WebLLM and WebGPU in a Web Worker. Under a gigabyte of GPU memory and a 4,096-token context.
This is not a video of a chatbot. It is a real 0.5B language model running on your own GPU — about 265 MB, once. It knows Parth's projects, skills and story; flip Web on and general questions can look things up on Wikipedia. The conversation never leaves this tab unless you ask it to.
A real language model, running on your GPU. It knows Parth's projects, skills and story, and loads only when you ask.
~265 MB one-time download · runs on your GPU · nothing you type leaves this tab
Qwen2.5 0.5B, 4-bit, through WebLLM and WebGPU in a Web Worker. Under a gigabyte of GPU memory and a 4,096-token context.
Nothing, once the model is cached. No analytics, no cookies, no chat storage. With Web on, only a general question is sent to Wikipedia.
Without WebGPU, the machine answers instantly from the same notes. The rest of the site never depends on it.
Four steps, no server in the loop. The only network requests are the model download and, if you turn it on, the Wikipedia lookups Web mode makes for you.
The site checks for WebGPU. If it is available, it starts a dedicated Web Worker and begins the one-time model download.
Qwen2.5-0.5B-Instruct-q4f16_1 is fetched from a public CDN, quantised to 4-bit, and loaded into GPU memory through WebLLM. The browser caches the weights.
Plain code works out what you're asking — which project, which detail, whether it follows on from the last answer — and hands the model the two or three fact cards that answer it. A small model writes best from a short, focused brief.
Tokens are generated on your GPU and streamed into the chat. Sources are shown as chips. Reload the page and the conversation is gone.
Small enough to load quickly, large enough to answer a question about the work.