What if your AI model could run directly in the browser, with no server calls, no tokens, no network latency? WebLLM makes this possible: it runs quantized language models directly in the browser, leveraging WebGPU, without a single token leaving the user's machine.

In this talk I'll share how I integrated it in Angular, starting from an experiment and ending up with a working game with NPC powered by a local LLM.

We'll explore together how WebLLM works, how I integrated it in Angular, and the strategies I used to work around its limits and get the most out of it.