把正在发生的 AI 拿在手里玩一玩：机器人、世界模型、3D、语音和科学 AI，也有投研工具、文档解析、智能体、代码、翻译、搜索和评测榜单。每张卡片点开就能试。
Get your hands on the AI that is happening right now: robots, world models, 3D, speech and AI for science, plus research tools, document parsing, agents, code, translation, search and leaderboards. Open any card and try it.
HuggingFace 在部分网络环境下打不开。加载失败时，卡片会改成「在新窗口打开」，可能需要国际网络。 HuggingFace is unreachable on some networks. If a demo cannot load, the card switches to an "open in a new window" link.
00
伙伴展台
Partner showcase
seeles.aiSEELESeeleAgent
伙伴展台 · SEELE
一句话，一个智能体：游戏、图片、视频、3D 与代码，端到端完成
One prompt. One agent. Games, images, videos, 3D assets, and code — built end to end.
一句话生成可玩的游戏One prompt to a playable game从一句提示词，生成可以玩的完整游戏。One prompt to a playable game.
一个标签页做完One place to think, build, ship在同一个云端工作区里生成图片、视频、3D 资产、游戏和网页项目，文件、预览、对话和运行都在一个标签页里。Generate images, video, 3D assets, game builds and web projects; files, preview, agent chat and runtime in one tab.
自研基础模型In-house foundation models由多模态基础模型 Seele02 和 Seele PEGA 世界模型驱动，后者用于实时、物理驱动的 3D 世界。Powered by Seele02, a native multimodal foundation model, and the Seele PEGA Worldmodel for real-time, physics-driven 3D worlds.
三种引擎，开箱即多人Three engines, multiplayer out of the box支持 Three.js、Unity 6 和 Unreal，在浏览器里构建。Three.js, Unity 6 and Unreal, built in the browser.
技能库A growing library of skills技能库不断扩充，每种模态都有对应技能；调用一个技能，智能体端到端完成。A growing library of skills, one for every modality; invoke one and the agent runs it end to end.
300 万+ 创作者3M+ creators截至 2026 年，受到 300 万以上创作者信赖，可以免费开始。Trusted by 3M+ creators as of 2026. Free to start.
数据来源：SEELE 官网 seeles.ai（首页，2026-10-18 查阅）。原文为英文，中文为直译。 Source: seeles.ai (home page), accessed 2026-10-18. Original text in English; Chinese is a literal translation.
伙伴近况
Partner updates
SEELEseeles.ai
LinkedIn
Introducing SEELE02-ultra. Create interactive 3D games from your ideas.
来源：官方账号公开发布。标题照录原文，另一种语言为直译。更新于 2026-10-18 19:00（北京时间）。 Source: public posts from official accounts. Titles are quoted as published; the other language is a literal translation. Updated 2026-10-18 19:00 (Beijing time).
A long-form robot learning tutorial from the University of Oxford and Hugging Face: classical control, reinforcement learning, imitation learning and generalist robot policies, with runnable code.
为什么推荐想系统看懂具身赛道的技术路线，这是一份免费、成体系的读物。Why it matters A free, structured read for anyone who wants to understand the technical roadmap of embodied AI.
NVIDIA's viewer for robot action data running the Cosmos3 world model, with three modes: predict future frames from actions, infer actions from frames, or predict the next actions as a policy.
为什么推荐一个界面讲清「世界模型怎样给机器人用」，适合给非技术同事演示。Why it matters One screen that explains how world models serve robots; good for showing non-technical colleagues.
需要浏览器开启 WebGL 与硬件加速。 Needs WebGL and hardware acceleration.
3D asset generation for embodied-AI simulation: describe an object in words, get a 3D model.
为什么推荐为机器人仿真批量造 3D 物体，是具身数据管线里容易被忽略的一环。Why it matters Mass-producing 3D objects for robot simulation, an easily overlooked link in the embodied data pipeline.
HorizonRobotics许可证 License · Apache-2.0点赞 Likes · 14主组 · 3D 3D
腾讯混元的通用 3D 世界重建：输入几张图片（可选相机位姿、深度等先验），一次前向输出点云、深度、法线和相机参数。
Tencent Hunyuan's universal 3D world reconstruction: feed in a few images (optionally with camera poses, depth and other priors) and get point clouds, depth, normals and camera parameters in one forward pass.
为什么推荐世界模型和 3D 重建正在合流，这个 demo 把两者放在一起看。Why it matters World models and 3D reconstruction are converging; this demo shows both at once.
NVIDIA's viewer for robot action data running the Cosmos3 world model, with three modes: predict future frames from actions, infer actions from frames, or predict the next actions as a policy.
为什么推荐一个界面讲清「世界模型怎样给机器人用」，适合给非技术同事演示。Why it matters One screen that explains how world models serve robots; good for showing non-technical colleagues.
需要浏览器开启 WebGL 与硬件加速。 Needs WebGL and hardware acceleration.
High-fidelity 3D generation from a single image; rotate it and export.
为什么推荐一张图到高精度 3D 资产，是 3D 内容生产成本下降最直观的证据。Why it matters One image to a high-fidelity 3D asset: the clearest evidence that 3D content is getting cheap to make.
3D asset generation for embodied-AI simulation: describe an object in words, get a 3D model.
为什么推荐为机器人仿真批量造 3D 物体，是具身数据管线里容易被忽略的一环。Why it matters Mass-producing 3D objects for robot simulation, an easily overlooked link in the embodied data pipeline.
腾讯混元的通用 3D 世界重建：输入几张图片（可选相机位姿、深度等先验），一次前向输出点云、深度、法线和相机参数。
Tencent Hunyuan's universal 3D world reconstruction: feed in a few images (optionally with camera poses, depth and other priors) and get point clouds, depth, normals and camera parameters in one forward pass.
为什么推荐世界模型和 3D 重建正在合流，这个 demo 把两者放在一起看。Why it matters World models and 3D reconstruction are converging; this demo shows both at once.
Speech synthesis with an 82M-parameter model, running locally in your browser on WebGPU.
为什么推荐语音合成完全在你的浏览器里跑，不上传任何文字，演示端侧隐私最直接。Why it matters Speech synthesis entirely in your browser, nothing uploaded: the most direct demo of on-device privacy.
Speech synthesis with an 82M-parameter model, running locally in your browser on WebGPU.
为什么推荐语音合成完全在你的浏览器里跑，不上传任何文字，演示端侧隐私最直接。Why it matters Speech synthesis entirely in your browser, nothing uploaded: the most direct demo of on-device privacy.
过去一年各机构在 HuggingFace 上发布的科学 AI 模型、数据集和应用热力图，可按生物、化学、物理、材料等领域筛选。
A heatmap of the open AI-for-science models, datasets and apps released on Hugging Face over the past year, filterable by biology, chemistry, physics, materials and more.
为什么推荐一眼看出谁在科学 AI 上最活跃，是做赛道扫描的现成工具。Why it matters See at a glance who is most active in AI for science; a ready-made tool for sector scans.
A vision-language model for charts: upload a chart and ask about trends, shares and exact values.
为什么推荐研报和招股书里的图表很多，这类模型能把图变回可核对的数字。Why it matters Reports and prospectuses are full of charts; models like this turn them back into numbers you can check.
DeepSeek-OCR 2 演示：把 PDF 或图片转成 Markdown，抽取文字、解析图表，还能按描述定位内容位置。
A DeepSeek-OCR 2 demo: convert PDFs or images to Markdown, extract text, parse figures, and locate content by description.
为什么推荐「先压缩成视觉 token 再读」的新思路，长文档处理成本可能因此下降。Why it matters A new approach that compresses pages into visual tokens before reading, which may cut the cost of long documents.
Berkeley's function calling leaderboard (BFCL), measuring how accurately LLMs call tools, now extended to multi-turn and agentic tasks.
为什么推荐智能体好不好用，关键在会不会正确调用工具，这张榜单专门量这一点。Why it matters Whether an agent is useful comes down to calling tools correctly; this board measures exactly that.
Hugging Face's official open LLM leaderboard, with 4,500+ models evaluated reproducibly on the same benchmarks; now archived, with historical rankings kept.
为什么推荐开源模型这两年怎么一步步追上来的，这里有最完整的记录。Why it matters The most complete record of how open models caught up over the past two years.
Berkeley's function calling leaderboard (BFCL), measuring how accurately LLMs call tools, now extended to multi-turn and agentic tasks.
为什么推荐智能体好不好用，关键在会不会正确调用工具，这张榜单专门量这一点。Why it matters Whether an agent is useful comes down to calling tools correctly; this board measures exactly that.
These are public projects from the HuggingFace community. All rights belong to their authors. Curated by the Kangpeng Institute.
项目运行在 HuggingFace 上，本站不收集你在其中输入的内容；请留意各项目自己的条款。点赞数和运行状态核实于 2026-10-19 15:27–15:37 +08:00（北京时间），之后可能变化。 Demos run on HuggingFace; this site does not collect what you enter in them. Each project has its own terms. Likes and status verified 2026-10-19 15:27–15:37 +08:00 (Beijing time) and may change.
