{"id":23766,"date":"2026-09-06T04:18:46","date_gmt":"2026-09-06T04:18:46","guid":{"rendered":"https:\/\/scannn.com\/running-llms-in-the-browser-with-three-js\/"},"modified":"2026-09-06T04:18:46","modified_gmt":"2026-09-06T04:18:46","slug":"running-llms-in-the-browser-with-three-js","status":"publish","type":"post","link":"https:\/\/scannn.com\/lv\/running-llms-in-the-browser-with-three-js\/","title":{"rendered":"Running LLMs in the Browser with Three.js"},"content":{"rendered":"\n
\n

TL;DR#<\/a><\/h2>\n

Three-LLM runs GPT-2, SmolLM2, Qwen, and Phi models locally in the browser by
\nturning their inference graphs into Three.js TSL compute shaders on WebGPU. Try
\nthe demo at https:\/\/three-llm.ben3d.ca<\/a>.<\/p>\n

I started this project to test two things: how far I could push the general
\ncompute capabilities of Three.js and WebGPU, and what today’s small language
\nmodels can do when they run in a browser. The models range from a 3-million
\nparameter story generator to a modern 800-million parameter chat model.<\/p>\n

Three.js is known as a rendering library, but its WebGPU renderer also exposes
\nstorage buffers, compute dispatches, workgroup memory, atomics, and GPU
\nreadback. A language model puts all of those capabilities to work. It needs
\nhundreds of ordered compute dispatches, large matrix-vector products,
\nreductions, persistent attention or recurrent state, and a tight loop between
\nthe GPU and JavaScript as each new token appears.<\/p>\n

I built Three-LLM<\/a> as both a test of
\nthat compute stack and a usable inference library. It loads ordinary Hugging
\nFace configurations, tokenizers, and SafeTensors checkpoints in the browser,
\nbuilds the model from reusable
\nThree.js Shading Language<\/a> (TSL) compute
\nkernels, and runs inference on the user’s GPU through WebGPU. It does not need
\na server-side inference runtime or model-specific WebAssembly binary.<\/p>\n

The package includes CPU reference implementations for validation and GPU
\nimplementations for actual use. The library currently recognizes GPT-2,
\nLlama-style, Gemma 3, Phi, and Qwen3.5 architectures. I also made a
\nquick chat application<\/a> so you can try five
\ncheckpoints without writing an application first:<\/p>\n