#WebGPU
8 related articles
Running LLMs in the Browser: WebGPU Hits 63 tokens/s, WASM Only 1.6
BEINGS chatroom benchmark: the same 360M model hits 62.9 tokens/s on WebGPU vs 1.6 on WASM—0.54 s vs 26.9 s per reply. Full lessons on Transformers.js, Web Workers, sampling, and bundler traps.
61KB HTML Delivers Real-Time WebGPU Ocean: Generated from a Single Claude Code Prompt
A developer used a single Claude Code prompt to generate a 61KB HTML file with a real-time WebGPU ocean: FFT waves, whitecaps, physical sky, in ~30 minutes, 260 FPS on an RTX 5070 Ti.
Zero-Token Decisions: EmbeddingGemma 2 + WebGPU for In-Browser Multimodal Classification
Decide Live uses fine-tuned EmbeddingGemma 2 with transformers.js and WebGPU to run zero-token multimodal classification entirely in-browser — no uploads, no API costs, offline-capable.
Talus: A 23M-Parameter Diffusion Model That Generates Game Terrain via Browser WebGPU
Talus is an open-source 23M-parameter diffusion model that generates game terrain heightmaps in ~3 seconds via browser WebGPU. Trained on a single RTX 5060, Apache-2.0 licensed.
EmbeddingGemma 2 In-Browser Image-Text Retrieval: Local Inference with WebGPU
EmbeddingGemma 2 maps images and text into a shared 768-dim vector space for retrieval. With the WebGPU inference library ruNNtime, semantic photo search runs entirely in the browser — no server needed.
Hugging Face Open-Sources the World's Fastest WebGPU Kernels for Local Browser AI
Hugging Face open-sources the world's fastest WebGPU kernels for local browser AI, covering 200+ ML operators with plans to upstream into Transformers.js and ONNX Runtime Web.
vgpu: A WebGPU TypeScript Library Designed for AI Agents
vgpu is a WebGPU TypeScript library built for AI agents, featuring typed shader imports, a minimal GPU-first API, and unified browser/Node/test suite execution.
Ternary Weight Compression: 27B Model Shrunk to 6GB, Runs Locally in the Browser
Ternary Bonsai 2 (27B) compresses a 27B model to under 6GB using ternary weights — one-ninth of FP16 — while claiming 98.2% performance retention and browser-local inference via WebGPU.