Releases·Tool Release
Hugging Face launches Fleet, browser-based WebGPU benchmarking with open-sourced kernels
Hugging Face released Fleet, a browser tool that runs WebGPU compute kernels from real AI workloads on users' GPUs and returns a personalized benchmark card. It also open-sourced hundreds of the kernels as a first step toward faster browser inference.
25 engagement2 sourcesTue, Sep 1, 2026, 04:00 PM
AI Overview
- Runs WebGPU compute kernels drawn from real AI workloads directly on the user's hardware, producing a personalized benchmark card per device
- Open-sources hundreds of WebGPU kernels, described as a first step toward making browser inference as fast as physically possible
- Contributed results show kernel performance across real hardware and help make the kernels faster
- Benchmark available at webgpu-kernels-fleet.hf.space with a blog post on huggingface.co
Entities
Related
Releases·Tool Release
Software Mansion releases runntime, a TypeGPU-based framework for running AI models on the user's GPU in WebGPU environments
Software Mansion released ruNNtime, a framework for running AI models on the user's own GPU via WebGPU, built on TypeGPU. Kernels are written in TypeGPU and compiled to WGSL at runtime, so no WASM binary is needed and the same code runs in the browser, React Native, and Electron. It offers a 'zoo' of one-function-per-task models for speech, vision, and NLP, positioning itself as an alternative to transformers.js for on-device web inference.
Software Mansion released ruNNtime, a framework for running AI models on the user's own GPU via WebGPU, built on TypeGPU. Kernels are written in TypeGPU and compiled to WGSL at runtime, so no WASM binary is needed and the same code runs in the browser, React Native, and Electron. It offers a 'zoo' of one-function-per-task models for speech, vision, and NLP, positioning itself as an alternative to transformers.js for on-device web inference.
37 engagement1 sourcegithubMon, Sep 14, 2026, 01:28 PM
Releases·Tool Release
LFM 2.5 230M achieves 1440 tok/s in-browser via custom WebGPU backend
A developer released a custom WebGPU backend for running LFM 2.5 230M in-browser, achieving 1400-1500 tok/s on an RTX 3090 and 400-500 tok/s on an M4. The backend supports Nvidia and Apple Silicon with hardware-specific kernel optimizations, and is portable to Electron/Tauri apps.
A developer released a custom WebGPU backend for running LFM 2.5 230M in-browser, achieving 1400-1500 tok/s on an RTX 3090 and 400-500 tok/s on an M4. The backend supports Nvidia and Apple Silicon with hardware-specific kernel optimizations, and is portable to Electron/Tauri apps.
56 engagement1 sourceredditSat, Jul 25, 2026, 05:14 PM
Products·launched
1-bit LLM inference in browser via WebGPU
A WebGPU-accelerated demo that runs a 1-bit quantized LLM entirely in the browser. It enables local, private text generation without server calls, targeting developers and researchers exploring edge inference.
A WebGPU-accelerated demo that runs a 1-bit quantized LLM entirely in the browser. It enables local, private text generation without server calls, targeting developers and researchers exploring edge inference.
22 engagement1 sourcehackernewsThu, Jul 16, 2026, 04:50 PM
Releases·Model Release
Muse Glimmer 30B runs in-browser via custom WebGPU kernels at ~25 tok/s on M4 Max
A developer demonstrated Muse Glimmer 30B running fully in-browser using custom WebGPU kernels, achieving ~25 tokens per second on an M4 Max—matching llama.cpp performance. This shows WebGPU's potential for local LLM inference without native backends.
A developer demonstrated Muse Glimmer 30B running fully in-browser using custom WebGPU kernels, achieving ~25 tokens per second on an M4 Max—matching llama.cpp performance. This shows WebGPU's potential for local LLM inference without native backends.
48 engagement1 sourceredditTue, Aug 11, 2026, 04:39 PM
Products·building
Code-rendered video generation using a local quantized LLM instead of a frontier cloud model
A recreation of the 'Opus 5.5 makes videos' pipeline (model writes HTML/Canvas scenes, headless browser renders frames, ffmpeg encodes) run entirely on a local NVFP4-quantized Qwen 3.8 27B with no API or GPU rental. The LLM writes deterministic time-based canvas scenes plus TTS narration and synthesized background music. It shows solo local-LLM users they can produce code-rendered explainer videos without frontier cloud models.
A recreation of the 'Opus 5.5 makes videos' pipeline (model writes HTML/Canvas scenes, headless browser renders frames, ffmpeg encodes) run entirely on a local NVFP4-quantized Qwen 3.8 27B with no API or GPU rental. The LLM writes deterministic time-based canvas scenes plus TTS narration and synthesized background music. It shows solo local-LLM users they can produce code-rendered explainer videos without frontier cloud models.
13 engagement1 sourceredditTue, Oct 6, 2026, 08:34 PM