[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-matrix-game-4-0-open-source-60fps":3,"topics-all":38,"news-related-bef7a773-36fb-406b-949d-8d728b97b242":57},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"bef7a773-36fb-406b-949d-8d728b97b242","Matrix-Game 4.0 开源:单卡 60FPS 的实时可玩游戏世界","昆仑万维 Skywork 开源 Matrix-Game 4.0:游戏逻辑留在可执行引擎,自回归视频模型只把低模画面渲染成高保真视觉,按键到显示延迟 112 毫秒,单张 RTX 5090 跑 480p 60FPS,代码、权重与三款游戏资产以 Apache-2.0 开源,路线从生成视频转向引擎分工。","把一个视频生成模型做成「能玩的游戏」,最难的不是画面,而是三件事:世界状态要持久、游戏规则要可执行、交互延迟要低到玩家感觉不到。Skywork AI(昆仑万维)的 Matrix-Game 系列一直在走这条路线,10 月 15 日放出的 Matrix-Game 4.0 给出了一个明确的技术答案:别让模型一个人扛全部——把游戏逻辑还给引擎,神经渲染只管画面。\n\n## 引擎管规则,渲染管画面\n\nMatrix-Game 4.0 的核心设计是把「游戏执行」和「视觉生成」拆开。一套轻量游戏引擎维护世界状态、执行交互规则和物理;一个多模态编码智能体负责把玩家的自然语言想法翻译成游戏代码与动态场景字幕;真正出画面的是一个基于自回归视频模型的渲染器,它把引擎渲染出的低多边形代理帧转化成高保真视觉,条件输入是动态字幕和记忆参考。\n\n渲染器用成对的低模\u002F高保真视频训练,再蒸馏成一个因果生成器,按四帧一个块往前走,每块只需一到两步去噪。这套引擎持续引导视觉生成的框架,也让玩家可以用自然语言创建和修改场景与玩法。\n\n## 单卡 60FPS 是什么概念\n\n数字来自官方论文与技术报告:在单张 NVIDIA RTX 5090 上,Matrix-Game 4.0 实现 480p、60FPS 的游戏画面,按键到显示延迟 112 毫秒;单步生成模式下稳态吞吐 83FPS,峰值显存 18.2GiB。固定提示词、配合 SageAttention 时,RTX 5080 到 67FPS、RTX 5070 Ti 到 58FPS。5090 路径用上了 FP8 线性层、融合内核、CUDA 图和可选的 SageAttention 后端,仓库还提供 A100 的单卡推理路径和浏览器在线 demo。\n\n首发放出三款场景:中世纪城堡、古埃及(均为第三人称射击)和城市赛车。代码、模型权重和低模游戏资产以 Apache-2.0 许可证开源,GitHub 仓库目前已有 2.7k star。\n\n## 与端到端路线的分岔\n\n此前世界模型做「可玩世界」,主流是在生成端卷速度——八月底开源的 ForgeWM 用一步去噪做到 72FPS。Matrix-Game 4.0 换了个方向:规则、物理、玩家输入全部留在可执行引擎里,模型只负责把低模画面变好看。好处很直接——游戏机制由引擎强制执行,不会像纯生成模型那样出现物理漂移,玩家输入也能得到确定性响应;代价同样清楚——能玩什么受低模资产和场景库限制,首发只有三款场景,分辨率停在 480p,5080\u002F5070 Ti 的 67\u002F58FPS 还要固定提示词才拿得到。\n\n从 2025 年 5 月的 1.0 到今天的 4.0,这个系列已经迭代四代,路线判断一以贯之:可玩性是工程问题,不只是模型问题。对做交互内容的人来说,值得盯的可能不是下一个更大的视频模型,而是延迟、状态持久和规则执行这套工程闭环怎么搭起来。\n\n参考:arXiv:2610.27999;github.com\u002FSkyworkAI\u002FMatrix-Game","https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.27999","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"499f4b56-819d-49a3-9609-33e775143b86","multimodal",{"id":19,"name":20,"slug":20,"description":14,"color":14},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",{"id":22,"name":23,"slug":23,"description":14,"color":14},"ebe5dcd1-46b1-4298-b8c2-8e0e2f456e56","video-generation",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"0a6b14c5-1978-4d97-87c7-ccd8a10376e9","en","Matrix-Game 4.0 open source: 60FPS playable worlds on one GPU","Skywork's Matrix-Game 4.0 keeps game logic in an engine and gives visuals to a causal video model: 112ms latency, 60FPS on RTX 5090, Apache-2.0 code.","Making a video generation model into a playable game is hard for three reasons: the world state must persist, the game rules must be executable, and interaction latency must stay low enough that players never notice it. Skywork AI's Matrix-Game series has been working this route, and Matrix-Game 4.0, released October 15, gives a clear technical answer: don't make one model carry everything — hand game logic back to an engine, and let neural rendering handle only the visuals.\n\n## An engine for rules, a renderer for pixels\n\nThe core design of Matrix-Game 4.0 splits game execution from visual generation. A lightweight game engine maintains world state and executes interaction rules and physics; a multimodal coding agent translates a player's natural-language ideas into game code and dynamic scene captions; the actual frames come from a renderer built on an autoregressive video model, which converts the engine's low-poly proxy frames into detailed visuals, conditioned on dynamic captions and memory references.\n\nThe renderer is trained on aligned low-poly and high-fidelity video pairs, then distilled into a causal generator that moves forward in four-frame chunks with one or two denoising steps each. This framework of continuous engine guidance also lets players create and modify scenes and gameplay in natural language.\n\n## What single-GPU 60FPS means\n\nThe numbers come from the official paper and technical report: on a single NVIDIA RTX 5090, Matrix-Game 4.0 delivers 480p gameplay at 60FPS with 112ms key-press-to-display latency; single-step generation reaches 83FPS steady-state throughput with 18.2GiB peak GPU memory. With fixed prompts and SageAttention, throughput reaches 67FPS on an RTX 5080 and 58FPS on an RTX 5070 Ti. The 5090 path uses FP8 linear layers, fused kernels, CUDA graphs and an optional SageAttention backend; the repository also ships a single-GPU inference path for the A100 and a browser demo.\n\nThree scenes ship at launch: Medieval Castle and Ancient Egypt (both third-person shooters) plus City Racing. Code, model weights and low-poly game assets are open-sourced under Apache-2.0, and the GitHub repo has collected 2.7k stars.\n\n## A fork from the end-to-end route\n\nMost earlier playable-world work chased speed on the generation side — in late August, ForgeWM reached 72FPS with one-step denoising. Matrix-Game 4.0 takes a different direction: rules, physics and player input all stay inside an executable engine, and the model only makes the low-poly frames look good. The benefit is direct — mechanics are engine-enforced, so there is none of the physics drift you get from purely generative models, and player input gets deterministic responses. The cost is equally clear: what you can play is bounded by the low-poly asset and scene library (three scenes at launch), resolution stays at 480p, and the 67\u002F58FPS figures on the 5080\u002F5070 Ti require fixed prompts.\n\nFrom version 1.0 in May 2025 to 4.0 today, the series has iterated four times with one consistent judgment: playability is an engineering problem, not only a model problem. For anyone building interactive content, the thing to watch may not be the next bigger video model, but how the engineering loop of latency, persistent state and rule execution gets assembled.\n\nRefs: arXiv:2610.27999; github.com\u002FSkyworkAI\u002FMatrix-Game","matrix-game-4-0-open-source-60fps","2026-10-21T13:09:32Z","2026-10-21T13:09:39.587052Z","2026-10-21T13:09:39.587062Z",true,"agent",96,[39,48],{"slug":40,"tag_slug":40,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":35,"created_at":46,"modified_at":47},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":49,"tag_slug":49,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":35,"created_at":55,"modified_at":56},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":58},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"7ddc323f-fc52-406a-b6df-79b7393e121b","高德开源 DreamX-Creator:7B 原生音视频生成,2K 输出","dreamx-creator-7b-native-audio-video","2026-09-01T13:10:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"6f9e9f94-9dcc-4c6c-b254-6c5d0fe8ed37","京东开源 JoyAI-Video-Edit:16B 多模态扩散 Transformer 把视频编辑推进「边播边改」实时流时代","jd-joyai-video-edit-realtime-diffusion","2026-08-10T00:00:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"bcdc10bc-2f08-4c39-8ffa-e7e34041c112","京东开源 JoyAI-Video-Edit:用 16B 多模态扩散 Transformer 把视频编辑推进「边播边改」实时流时代","jd-joyai-video-edit-real-time-streaming","2026-08-05T03:00:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"6e3002da-c1fd-4a6d-b903-4f65b976dd04","MiniMax H3 首个商用落点：美图 RoboNeo 接入背后,通用多模态模型的\"可编辑性\"才刚开始被检验","roboneo-minimax-h3-multimodal-editing","2026-08-03T18:02:02+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"9f82c248-0592-421f-9fd8-ebd2100dcaf5","VideoChat3 全开源 4B 视频 MLLM 一次打通四种能力,I3D-ViT 把时空 token 砍掉 16×","videochat3-4b-mllm","2026-07-15T02:00:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"3851a096-37d6-4a45-bfb7-57b2fd65d992","京东开源 JoyAI-Echo：5 分钟长视频生成首次解决「跨镜头一致性」难题，DMD 蒸馏跑出 7.5× 加速","joyai-echo-jd-5-min-cross-shot-dmd-7-5x","2026-06-12T02:01:00+00:00"]