Tag
Fable 5.2 generated a fully procedural, flyable 3D Atlantis in a single 96 KB HTML file using Claude Code — including a WebGPU renderer, crystal refraction shaders, procedural buildings, collision, and a synthesized chiptune soundtrack — hitting ~2.5 ms frame times at 1440p on an RTX 5070 Ti.
An explorable O'Neill cylinder 3D world built entirely in code as a single 1064 KB HTML file using WebGPU—no images, models, assets, or libraries—created with AI models in AGNT.
Reverb-TTS is an open-source zero-shot voice cloning TTS model that runs fully locally in the browser on WebGPU via ruNNtime, achieving about 17x realtime on an M4 Pro with no server round-trips.
Show HN: Weight-Space Playground demos an LLM editing just 0.016% of another model's weights (131,072 of ~800M) to cure repetition loops in Qwen3.5-0.8B, cutting token-limit hits from 93% to 33.1% and boosting MATH-500 accuracy from 7.9% to 48.1% — all viewable and runnable in-browser via WebGPU.
A developer added Inworld's 64M-parameter voice isolator v1 to their caption app, running fully in-browser on a handwritten WebGPU engine with fp16 weights, cleaning a 20-second clip in about 6 seconds with no uploads. The demo is live at cuecaps.com/studio, with noise removal free and export behind a paid plan.
Voidstruck is a vibe-coded procedural space exploration game featuring 150 million star systems packed into a single 299 kB HTML file that runs in the browser with WebGPU, no assets or libraries, and live-synthesized music.
The creator built CueCaps Studio, a browser-based video captioning tool running Whisper base via transformers.js with WebGPU/WASM fallback, plus an 11M-param BERT tagger distilled from LFM3-Omni-2B that handles emphasis, line breaks, and punctuation.
A progress report on pushing Three.js native rendering to AAA-game-engine lighting quality, detailing a pipeline with G-buffer prepass, baked GI caches (lightmaps, light probe grids, reflection probes), SSR/SSGI, TRAA/TAAU, and PMREM — with a stage-by-stage breakdown of what is already in Three.js, in review, or still to build.
A fully open-source Three.js tech demo built in one shot with Claude Sonnet, rendering millions of GPU-generated Gaussian splats with zero external assets, foldable world math via Jacobian covariance transport, and five procedurally generated realms shipped as a single 1.83 MB HTML file.
Ava is an in-browser, fully local AI avatar powered by the LFM3-Omni model running via WebGPU, available as a Hugging Face Space demo.
A humorously named Gemma 4 E2B fine-tune called CRAP (Compiler Replacement for Assembling Programs) that emits working x86_64 ELF binaries directly from text prompts, skipping compilers and assemblers, now at v1.1 with a browser-based WebGPU demo.
LiquidAI 的 LFM3-Omni-2B 成为首个可在浏览器中通过 WebGPU 100% 本地运行的全模态模型,无需独立 ASR/TTS,在 M4 Max 上实现约 280 token/s、90ms 首 token 延迟。
Xenova 宣布突破:首个 100% 本地在浏览器中运行的真正全模态模型 LFM3-Omni-2B,借助自定义 WebGPU kernels 实现约 280 tok/s,支持视觉、音频和文本输入并生成交错的文本与音频输出。
Liquid AI released LFM3 on Hugging Face, a ~2B shared-backbone model family spanning text, vision, audio, omni, and decision models, capable of real-time local video captioning in the browser via WebGPU with Transformers.js.
LFM3-VL-2B 视觉语言模型可在浏览器中通过 WebGPU 本地实时进行视频字幕生成,LiquidAI 还发布了对应的 Hugging Face 模型集合与在线演示。
slughorn 是一个现代 C++20 GPU 矢量图形库,实现 Slug 渲染技术,支持从 SVG、FreeType、Skia 等导入或通过 Canvas API 直接绘制,并上传至 OpenGL、Vulkan、WebGPU、DirectX 等后端,构建时一次性完成栅格化以减少运行时开销。
作者微调了一个 360M 参数的 Qwen2.5-Coder-0.5B 模型,使其像真正的终端一样预测屏幕变化,可以模拟 vim、tmux、snake 等应用,展示了小型世界模型在字符网格界面中维持一致状态的可行性。
Odin is a browser-based debugging search engine that parses a repo and ranks the exact file (93%) and function (79.4%) to fix for a given bug, using Tree-sitter via WebAssembly and sub-1B local models via WebGPU, matching Claude 4.5 agent accuracy at a fraction of the cost on SWE-bench Lite.
作者在博客中演示了如何完全在浏览器端运行 RAG:用静态嵌入检索段落,再通过 transformers.js 在 WebGPU/WASM 上运行小型模型(如 Qwen3-0.6B、SmolLM2-360M)生成两句话摘要,全程数据不出本地。文章还评估了这些小模型的忠实度局限。
作者将上周的 MoE 专家按需从磁盘流式加载的方案接入 llama.cpp 的 WebGPU 后端,使 80B MoE 模型(Qwen3-Next,48GB GGUF)也能在 Chrome 标签页中运行,无需为每个模型手写引擎。该方案新增了 ggml-webgpu 的缓冲类型、slot 映射的 MUL_MAT_ID 着色器和驻留逻辑,实测在 M4 Pro MacBook 上以 2.9 tok/s 解码 80B 模型,较小模型可达 22 tok/s。