MicroLLM Lab:浏览器里跑 / 测 7 个小模型 MicroLLM Lab: run and benchmark 7 tiny LLMs in the browser
在浏览器里试跑小模型。 Try small language models in your browser.
在浏览器里试跑小模型。 Try small language models in your browser.
HN 首页 https://stateofutopia.com/experiments/microllmlab/ — 浏览器 WebGPU(可回退 WASM/CPU)加载 Q4 小模型,IndexedDB 缓存,免账号;有 Chat / Benchmarks / Compare 与可分享成绩证书。README 列出 PetitGPT / SmolLM2 / MiniMind2 / GPT-2 等权重许可证多为 Apache-2.0 或 MIT。 Chat / Benchmarks / Compare plus a shareable certificate. the README lists weight licenses mostly Apache-2.0 or MIT (PetitGPT / SmolLM2 / MiniMind2 / GPT-2, …).
在浏览器里生成 CSS/HTML 的小工具。 In-browser tools that generate CSS or HTML.
该类型下收录的公开 HTML 页面。 Public HTML pages listed under this type.
为什么推荐Why we recommend it 收录理由:立刻可试的本机小模型实验室,客观 suite 比聊天皮更有用。边界:仓库 SPDX 为 NOASSERTION,权重许可证以 README 表为准;无 WebGPU 时会慢很多(作者页也提示);tok/s 与准确率是本机跑出来的,本站未代跑全套。预览自托管。 Listing rationale: a tryable on-device small-model lab; the objective suite is more useful than a chat skin alone. Boundary: repo SPDX is NOASSERTION — weight licenses follow the README table; without WebGPU it is much slower (the page says so); tok/s and accuracy are measured on your machine — this site did not run the full suite.
主源是 HN 49882781 与公开实验页 / 仓库。 (item 49882781,2026-09-28T18:58:53Z / 约 2026-09-29 02:58 CST,Algolia 核对 116pts) 仓库 robss2020/microllm-lab 于 2026-09-29 gh api:8★、license NOASSERTION; 公开页与仓库 HTTP 200。 预览为 Playwright 实拍实验页裁 3:2 WebP。Primary sources are HN 49882781 and the public lab page / repo. Preview is self-hosted. (2026-09-28T18:58:53Z / about 2026-09-29 02:58 CST, 116pts on Algolia at check time) HN front page item 49882781 at https://stateofutopia.com/experiments/microllmlab/ — load Q4 small models in the browser over WebGPU (WASM/CPU fallback), cache in IndexedDB, no account; Repo robss2020/microllm-lab on the 2026-09-29 gh api: 8★, license NOASSERTION; Public page and repo HTTP 200. Preview is a Playwright capture of the lab page, cropped to 3:2 WebP.