{"@context":"https://schema.org","@type":"NewsArticle","generatedAt":"2026-08-11T09:21:12.743Z","headline":"Cursor 开源 Mixture-of-Kittens （MoK）：面向 GB300 NVL72 机架的确定性 MoE 训练 Megakernel","description":"Cursor Research 开源了其 Composer 模型背后的 MoE 训练 megakernel--Mixture-of-Kittens （MoK），将全部通信与计算融合进单一确定性内核，较最强公开基线最高提速 2.37 倍，并已在数万 GPU 上支撑 Composer 训练。","url":"https://www.aioga.com/news/cmsf13bs51dcsro2expu0v1pt/","mainEntityOfPage":"https://www.aioga.com/news/cmsf13bs51dcsro2expu0v1pt/","datePublished":"2026-08-04T18:38:41.000Z","dateModified":"2026-08-04T18:38:41.000Z","inLanguage":"zh-CN","publisher":{"@type":"NewsMediaOrganization","name":"Aioga","url":"https://www.aioga.com"},"citation":["https://www.marktechpost.com/2026/08/04/cursor-open-sources-mixture-of-kittens-mok-a-deterministic-moe-training-megakernel-for-gb300-nvl72-racks","https://aihot.virxact.com/items/cmsf13bs51dcsro2expu0v1pt"],"canonicalUrl":"https://www.aioga.com/news/cmsf13bs51dcsro2expu0v1pt/","directAnswer":{"@type":"Answer","text":"Aioga 编辑摘要：Cursor Research 开源了其 Composer 模型背后的 MoE 训练 megakernel--Mixture-of-Kittens Aioga 将其归入「产品更新」方向，重点关注它对真实使用和行业竞争的影响。","url":"https://www.aioga.com/news/cmsf13bs51dcsro2expu0v1pt/","dateCreated":"2026-08-04T18:38:41.000Z","author":{"@type":"Organization","@id":"https://www.aioga.com/authors/aioga-editorial/#editorial-team","name":"Aioga Editorial Team","url":"https://www.aioga.com/authors/aioga-editorial/"}},"evidence":[{"@type":"CreativeWork","name":"marktechpost.com source article","url":"https://www.marktechpost.com/2026/08/04/cursor-open-sources-mixture-of-kittens-mok-a-deterministic-moe-training-megakernel-for-gb300-nvl72-racks","datePublished":"2026-08-04T18:38:41.000Z","provider":{"@type":"Organization","name":"marktechpost.com","url":"https://www.marktechpost.com/2026/08/04/cursor-open-sources-mixture-of-kittens-mok-a-deterministic-moe-training-megakernel-for-gb300-nvl72-racks"}},{"@type":"CreativeWork","name":"AIHot archive record","url":"https://aihot.virxact.com/items/cmsf13bs51dcsro2expu0v1pt","datePublished":"2026-08-04T18:38:41.000Z","provider":{"@type":"Organization","name":"AIHot","url":"https://aihot.virxact.com/items/cmsf13bs51dcsro2expu0v1pt"}}],"aggregationSource":"MarkTechPost（RSS）","originalPublisher":{"name":"marktechpost.com","url":"https://www.marktechpost.com/2026/08/04/cursor-open-sources-mixture-of-kittens-mok-a-deterministic-moe-training-megakernel-for-gb300-nvl72-racks"},"geoDeepAnswer":null,"article":{"id":"cmsf13bs51dcsro2expu0v1pt","slug":"cmsf13bs51dcsro2expu0v1pt","url":"https://www.aioga.com/news/cmsf13bs51dcsro2expu0v1pt/","title":"Cursor 开源 Mixture-of-Kittens （MoK）：面向 GB300 NVL72 机架的确定性 MoE 训练 Megakernel","title_en":"Cursor Open-Sources Mixture-of-Kittens （MoK）： A Deterministic MoE Training Megakernel for GB300 NVL72 Racks","summary":"Cursor Research 开源了其 Composer 模型背后的 MoE 训练 megakernel--Mixture-of-Kittens （MoK），将全部通信与计算融合进单一确定性内核，较最强公开基线最高提速 2.37 倍，并已在数万 GPU 上支撑 Composer 训练。","source":"MarkTechPost（RSS）","sourceUrl":"https://www.marktechpost.com/2026/08/04/cursor-open-sources-mixture-of-kittens-mok-a-deterministic-moe-training-megakernel-for-gb300-nvl72-racks","aiHotUrl":"https://aihot.virxact.com/items/cmsf13bs51dcsro2expu0v1pt","publishedAt":"2026-08-04T18:38:41.000Z","category":"产品更新","score":30,"selected":false,"articleBody":["Cursor Research has open-sourced Mixture-of-Kittens (MoK)：https://cursor.com/blog/mixture-of-kittens, the mixture-of-experts training megakernel behind its Composer：https://cursor.com/blog/composer-2 models. MoK fuses every MoE communication and computation step into a single deterministic kernel. Cursor team reports up to 2.37x higher throughput than the strongest public baseline. It already powers Composer training across tens of thousands of GPUs.","Yes, but the hardware floor is high. MoK is on GitHub：https://github.com/cursor/mixture-of-kittens under Apache-2.0. It requires NVIDIA Blackwell SM100 or SM103 GPUs, which means GB200 NVL72 or GB300 NVL72 racks. It also needs Python 3.12+, PyTorch 2.10+, and CUDA toolkit 13.0+. Inter-GPU buffers rely on PyTorch symmetric memory.","That limits realistic adopters to organizations that own or rent NVL72 capacity. Frontier labs, funded model startups, GPU neoclouds, and national computing centers fit. Single-node teams and 8-GPU shops do not.","Applications are narrow but high-value. They include pretraining and post-training of DeepSeek-V3-style MoE models. Determinism also makes it useful for on-policy RL post-training and internal ablations. Relevant industries are AI model development, cloud GPU infrastructure, code-generation tooling, and quantitative research.","Cursor’s earlier work covered the compute side. The research team wrote its own MXFP8 and NVFP4 training kernels：https://cursor.com/blog/kernels and a ‘warp decode’：https://cursor.com/blog/warp-decode path for MoE inference. Those assumed inter-GPU communication was handled separately.","In production, communication became the limiting factor. The MoE layer can consume more than half of end-to-end training time. Moving to GB300 NVL72s changed the problem again. A rack is 72 GPUs inside one NVLink domain, which allows fine-grained overlap. But the integrated Grace CPUs are slow relative to the GPUs. CPU-GPU synchronization therefore has to be minimized aggressively.","MoK is built as a megakernel：https://hazyresearch.stanford.edu/blog/2025-05-27-no-bubbles and is fully deterministic. It supports BF16 and MXFP8 precision modes. Scheduling runs through Blackwell’s Cluster Launch Control, so inter-rack RDMA does not serialize behind it. Router weight gradients use a SonicMoE：https://arxiv.org/abs/2512.14080-style calculation fused into the SwiGLU backward.","Layer benchmarks ran in a single NVL72 rack at EP degree 64. Each GPU held 2,048 tokens before routing. Baselines were NCCL+PyTorch, DeepEP+PyTorch, DeepEP+TransformerEngine, and HybridEP+Megatron. Shapes covered Kimi K2.7 Code, GLM-5.2, Qwen3.5-397B-A17B, and DeepSeek-V4-Pro.","Against the fastest baseline, MoK is up to 2.37x faster for MXFP8 forward. The other figures are 1.78x MXFP8 backward, 1.92x BF16 forward, and 1.58x BF16 backward. End-to-end testing used 512 GPUs across several GB300 NVL72 racks. Tokens per second per GPU rose from 760.9 to 1,070.2, a 1.41x gain.","Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect with us ：https://forms.gle/wbash1wF6efRj8G58","Asif Razzaq is the CEO of Marktechpost Media Inc.. As a visionary entrepreneur and engineer, Asif is committed to harnessing the potential of Artificial Intelligence for social good. His most recent endeavor is the launch of an Artificial Intelligence Media Platform, Marktechpost, which stands out for its in-depth coverage of machine learning and deep learning news that is both technically sound and easily understandable by a wide audience. The platform boasts of over 2 million monthly views, illustrating its popularity among audiences.","[FREE GUIDE] Securing AI Agents, MCP Servers & LLM Apps ：https://pxllnk.co/lxn88m"],"articleImages":[{"sourceUrl":"https://www.marktechpost.com/wp-content/uploads/2026/08/moe-mxfp8-forward-1.png","alt":"","afterParagraph":6,"url":"/media/articles/cmsf13bs51dcsro2expu0v1pt/ad730be10899b3f8.webp"},{"sourceUrl":"https://www.marktechpost.com/wp-content/uploads/2019/06/Screen-Shot-2021-09-14-at-9.02.24-AM-300x300.png","alt":"","afterParagraph":9,"url":"/media/articles/cmsf13bs51dcsro2expu0v1pt/787a6d54564e8e19.webp"},{"sourceUrl":"https://www.marktechpost.com/wp-content/uploads/2026/08/high-level-description-a-monochrome-cybe_2rr6n78EUGGOOwTBaf3UzA_gSBnpjrRS5CwRLN0TlRb5g_cover_2k-100x70.png","alt":"Building an Advanced AI Skill Security Auditing Pipeline with NVIDIA SkillSpector, LangGraph, YARA Rules, SARIF, and CI Policy Gates","afterParagraph":10,"url":"/media/articles/cmsf13bs51dcsro2expu0v1pt/542fa098b88910cd.png"},{"sourceUrl":"https://www.marktechpost.com/wp-content/uploads/2026/08/blog6176-8-100x70.png","alt":"Y Combinator Open-Sources QM: An MIT-Licensed Multiplayer Agent Harness That Runs In Slack And The Web","afterParagraph":10,"url":"/media/articles/cmsf13bs51dcsro2expu0v1pt/93f58f09d44c5dc2.webp"},{"sourceUrl":"https://www.marktechpost.com/wp-content/uploads/2026/08/high-level-description-a-tech-news-cover_ZCsPPdgeWVyT_N_w-0YNbw_80NcwxTETLOSb0ADNIkbhA_cover_2k-100x70.png","alt":"Genspark Open Sources GenOffice: A Free, Ad-Free AI Office Suite for macOS and Windows with Docs, Sheets, Slides, PDF","afterParagraph":10,"url":"/media/articles/cmsf13bs51dcsro2expu0v1pt/6a2da86063b41ba7.webp"}],"mediaStatus":"ok","articleBodyZh":["Cursor Research 已开源 Mixture-of-Kittens (MoK)：https://cursor.com/blog/mixture-of-kittens，这是其 Composer 背后的混合专家训练超核：https://cursor.com/blog/composer-2 模型。MoK 将每个 MoE 的通信和计算步骤融合为一个确定性核。Cursor 团队报告称，其吞吐量比最强的公开基线高出最多 2.37 倍。它已经为数万 GPU 的 Composer 训练提供了支持。","是的，但硬件门槛很高。MoK 在 GitHub：https://github.com/cursor/mixture-of-kittens 上以 Apache-2.0 协议开源。它需要 NVIDIA Blackwell SM100 或 SM103 GPU，这意味着 GB200 NVL72 或 GB300 NVL72 机架。此外，还需要 Python 3.12+、PyTorch 2.10+ 和 CUDA 工具包 13.0+。GPU 之间的缓冲区依赖于 PyTorch 对称内存。","这限制了现实中的采用者，仅限拥有或租用 NVL72 容量的组织。前沿实验室、获得资金的模型初创公司、GPU 新云以及国家计算中心符合条件。单节点团队和 8-GPU 小型团队则不适用。","应用范围狭窄但价值高。包括 DeepSeek-V3 风格 MoE 模型的预训练和后训练。确定性也使其在策略内 RL 后训练和内部消融实验中有用。相关行业包括 AI 模型开发、云 GPU 基础设施、代码生成工具以及量化研究。","Cursor 早期的工作覆盖了计算方面。研究团队编写了自己的 MXFP8 和 NVFP4 训练核心：https://cursor.com/blog/kernels，以及用于 MoE 推理的“warp decode”路径：https://cursor.com/blog/warp-decode。这些假设 GPU 之间的通信是单独处理的。","在生产中，通信成为限制因素。MoE 层可能消耗超过端到端训练时间的一半。迁移到 GB300 NVL72 改变了问题。一机架内有 72 个 GPU，位于同一 NVLink 域内，这允许细粒度的重叠。但与 GPU 相比，集成的 Grace CPU 速度较慢。因此必须积极最小化 CPU 与 GPU 的同步。","MoK 构建为一个巨内核（megakernel）：https://hazyresearch.stanford.edu/blog/2025-05-27-no-bubbles，并且是完全确定性的。它支持 BF16 和 MXFP8 精度模式。调度通过 Blackwell 的 Cluster Launch Control 运行，因此机架间的 RDMA 不会在其后排队。路由器权重梯度使用 SonicMoE：https://arxiv.org/abs/2512.14080 风格的计算，融合到 SwiGLU 的反向传播中。","层基准测试在单个 NVL72 机架上以 EP 等级 64 运行。每个 GPU 在路由之前持有 2,048 个令牌。基线包括 NCCL+PyTorch、DeepEP+PyTorch、DeepEP+TransformerEngine 和 HybridEP+Megatron。覆盖的模型包括 Kimi K2.7 代码、GLM-5.2、Qwen3.5-397B-A17B 和 DeepSeek-V4-Pro。","相比最快的基准，MoK 的 MXFP8 正向运算速度最多提高 2.37 倍。其他数据为 MXFP8 反向运算 1.78 倍、BF16 正向运算 1.92 倍、BF16 反向运算 1.58 倍。端到端测试使用了跨多个 GB300 NVL72 机架的 512 个 GPU。每个 GPU 的每秒 token 数从 760.9 上升到 1,070.2，提升 1.41 倍。","需要与我们合作推广您的 GitHub 仓库、Hugging Face 页面、产品发布或网络研讨会等吗？请联系我们： https://forms.gle/wbash1wF6efRj8G58","Asif Razzaq 是 Marktechpost Media Inc. 的 CEO。作为一位有远见的企业家和工程师，Asif 致力于利用人工智能的潜力来造福社会。他最近的项目是推出人工智能媒体平台 Marktechpost，该平台以深入报道机器学习和深度学习新闻而闻名，内容既具有技术性，又容易被广大读者理解。该平台每月访问量超过 200 万，显示出其在受众中的受欢迎程度。","[免费指南] 保护 AI 代理、MCP 服务器及 LLM 应用：https://pxllnk.co/lxn88m"],"translationStatus":"translated","bodyOrigin":"source-page","editorial":{"summary":"Aioga 编辑摘要：Cursor Research 开源了其 Composer 模型背后的 MoE 训练 megakernel--Mixture-of-Kittens Aioga 将其归入「产品更新」方向，重点关注它对真实使用和行业竞争的影响。","background":"背景分析：模型与研究类动态需要结合能力边界、开放方式、成本、可用性和真实任务表现判断，单项指标领先不等于已经形成稳定采用。","viewpoint":"Aioga 判断：这条动态更适合作为行业观察信号，当前信息足以建立线索，但不足以推导长期结论。","implications":"影响分析：对相关团队而言，短期应先核对来源、可用范围和实际成本，再判断是否值得接入或跟进。","nextStep":"后续观察：继续观察官方文档、实际可用性、价格变化、开发者反馈和竞品回应。","evidenceRefs":["title","summary","articleBody"],"confidence":"medium","status":"published","aiGenerated":false,"autoApproved":true,"generatedBy":"rule-safe-fallback","generatedAt":"2026-08-11T09:23:27.681Z","sourceHash":"c1eea9a3d24a1e22","validation":{"passed":true,"mode":"rule-safe-fallback","checks":["schema","length","source-attribution","no-html"]}},"tags":["产品更新","MarkTechPost（RSS）"],"translations":{"zh-CN":{"title":"Cursor 开源 Mixture-of-Kittens （MoK）：面向 GB300 NVL72 机架的确定性 MoE 训练 Megakernel","summary":"Cursor Research 开源了其 Composer 模型背后的 MoE 训练 megakernel--Mixture-of-Kittens （MoK），将全部通信与计算融合进单一确定性内核，较最强公开基线最高提速 2.37 倍，并已在数万 GPU 上支撑 Composer 训练。","category":"产品更新","source":"marktechpost.com","aggregationSource":"MarkTechPost（RSS）","pageTitle":"Cursor 开源 Mixture-of-Kittens （MoK）：面向 GB300 NVL72 机架的确定性 MoE 训练 Megakernel - Aioga AI资讯","description":"Cursor Research 开源了其 Composer 模型背后的 MoE 训练 megakernel--Mixture-of-Kittens （MoK），将全部通信与计算融合进单一确定性内核，较最强公开基线最高提速 2.37 倍，并已在数万 GPU 上支撑 Composer 训练。","url":"https://www.aioga.com/news/cmsf13bs51dcsro2expu0v1pt/","articleBody":["Cursor Research 已开源 Mixture-of-Kittens (MoK)：https://cursor.com/blog/mixture-of-kittens，这是其 Composer 背后的混合专家训练超核：https://cursor.com/blog/composer-2 模型。MoK 将每个 MoE 的通信和计算步骤融合为一个确定性核。Cursor 团队报告称，其吞吐量比最强的公开基线高出最多 2.37 倍。它已经为数万 GPU 的 Composer 训练提供了支持。","是的，但硬件门槛很高。MoK 在 GitHub：https://github.com/cursor/mixture-of-kittens 上以 Apache-2.0 协议开源。它需要 NVIDIA Blackwell SM100 或 SM103 GPU，这意味着 GB200 NVL72 或 GB300 NVL72 机架。此外，还需要 Python 3.12+、PyTorch 2.10+ 和 CUDA 工具包 13.0+。GPU 之间的缓冲区依赖于 PyTorch 对称内存。","这限制了现实中的采用者，仅限拥有或租用 NVL72 容量的组织。前沿实验室、获得资金的模型初创公司、GPU 新云以及国家计算中心符合条件。单节点团队和 8-GPU 小型团队则不适用。","应用范围狭窄但价值高。包括 DeepSeek-V3 风格 MoE 模型的预训练和后训练。确定性也使其在策略内 RL 后训练和内部消融实验中有用。相关行业包括 AI 模型开发、云 GPU 基础设施、代码生成工具以及量化研究。","Cursor 早期的工作覆盖了计算方面。研究团队编写了自己的 MXFP8 和 NVFP4 训练核心：https://cursor.com/blog/kernels，以及用于 MoE 推理的“warp decode”路径：https://cursor.com/blog/warp-decode。这些假设 GPU 之间的通信是单独处理的。","在生产中，通信成为限制因素。MoE 层可能消耗超过端到端训练时间的一半。迁移到 GB300 NVL72 改变了问题。一机架内有 72 个 GPU，位于同一 NVLink 域内，这允许细粒度的重叠。但与 GPU 相比，集成的 Grace CPU 速度较慢。因此必须积极最小化 CPU 与 GPU 的同步。","MoK 构建为一个巨内核（megakernel）：https://hazyresearch.stanford.edu/blog/2025-05-27-no-bubbles，并且是完全确定性的。它支持 BF16 和 MXFP8 精度模式。调度通过 Blackwell 的 Cluster Launch Control 运行，因此机架间的 RDMA 不会在其后排队。路由器权重梯度使用 SonicMoE：https://arxiv.org/abs/2512.14080 风格的计算，融合到 SwiGLU 的反向传播中。","层基准测试在单个 NVL72 机架上以 EP 等级 64 运行。每个 GPU 在路由之前持有 2,048 个令牌。基线包括 NCCL+PyTorch、DeepEP+PyTorch、DeepEP+TransformerEngine 和 HybridEP+Megatron。覆盖的模型包括 Kimi K2.7 代码、GLM-5.2、Qwen3.5-397B-A17B 和 DeepSeek-V4-Pro。","相比最快的基准，MoK 的 MXFP8 正向运算速度最多提高 2.37 倍。其他数据为 MXFP8 反向运算 1.78 倍、BF16 正向运算 1.92 倍、BF16 反向运算 1.58 倍。端到端测试使用了跨多个 GB300 NVL72 机架的 512 个 GPU。每个 GPU 的每秒 token 数从 760.9 上升到 1,070.2，提升 1.41 倍。","需要与我们合作推广您的 GitHub 仓库、Hugging Face 页面、产品发布或网络研讨会等吗？请联系我们： https://forms.gle/wbash1wF6efRj8G58","Asif Razzaq 是 Marktechpost Media Inc. 的 CEO。作为一位有远见的企业家和工程师，Asif 致力于利用人工智能的潜力来造福社会。他最近的项目是推出人工智能媒体平台 Marktechpost，该平台以深入报道机器学习和深度学习新闻而闻名，内容既具有技术性，又容易被广大读者理解。该平台每月访问量超过 200 万，显示出其在受众中的受欢迎程度。","[免费指南] 保护 AI 代理、MCP 服务器及 LLM 应用：https://pxllnk.co/lxn88m"]},"en":{"title":"Cursor Open Source Mixture-of-Kittens (MoK): Deterministic MoE training Megakernel for GB300 NVL72 racks","summary":"Cursor Research has open-sourced the MoE training megakernel behind its Composer model—Mixture-of-Kittens (MoK), fusing all communication and computation into a single deterministic kernel, achieving speeds up to 2.37 times faster than the strongest public baseline, and already supports Composer training on tens of thousands of GPUs.","category":"Products","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"Cursor Open Source Mixture-of-Kittens (MoK): Deterministic MoE training Megakernel for GB300 NVL72 racks - Aioga AI News","description":"Cursor Research has open-sourced the MoE training megakernel behind its Composer model—Mixture-of-Kittens (MoK), fusing all communication and computation into a single deterministi...","url":"https://www.aioga.com/en/news/cmsf13bs51dcsro2expu0v1pt/","contentTranslated":true,"sourceHash":"c9ff0d876aa8ba41","translatedAt":"2026-08-04T20:25:24.216Z"},"ja":{"title":"Cursor Open Source Mixture-of-Kittens (MoK):GB300 NVL72ラック用の決定的MoEトレーニングメガカーネル","summary":"Cursor Researchは、Composerモデルの背後にあるMoEトレーニングメガカーネルをオープンソース化しました—Mixture-of-Kittens(MoK)。すべての通信と計算を単一の決定論的カーネルに融合させ、最強の公開ベースラインの最大2.37倍の速度を実現し、すでに数万台のGPUでComposerトレーニングをサポートしています。","category":"製品更新","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"Cursor Open Source Mixture-of-Kittens (MoK):GB300 NVL72ラック用の決定的MoEトレーニングメガカーネル - Aioga AIニュース","description":"Cursor Researchは、Composerモデルの背後にあるMoEトレーニングメガカーネルをオープンソース化しました—Mixture-of-Kittens(MoK)。すべての通信と計算を単一の決定論的カーネルに融合させ、最強の公開ベースラインの最大2.37倍の速度を実現し、すでに数万台のGPUでComposerトレーニングをサポートしています。","url":"https://www.aioga.com/ja/news/cmsf13bs51dcsro2expu0v1pt/","contentTranslated":true,"sourceHash":"c9ff0d876aa8ba41","translatedAt":"2026-08-04T20:25:22.837Z"},"ko":{"title":"Cursor Open Source Mixture-of-Kittens (MoK): GB300 NVL72 랙을 위한 결정적 MoE 훈련 메가커널","summary":"Cursor Research는 Composer 모델인 Mixture-of-Kittens(MoK) 뒤에 있는 MoE 훈련 메가커널을 오픈소스로 공개했으며, 모든 통신과 계산을 하나의 결정론적 커널로 융합하여 가장 강력한 공개 기준선보다 최대 2.37배 빠른 속도를 달성하고, 이미 수만 개의 GPU에서 Composer 훈련을 지원하고 있습니다.","category":"제품 업데이트","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"Cursor Open Source Mixture-of-Kittens (MoK): GB300 NVL72 랙을 위한 결정적 MoE 훈련 메가커널 - Aioga AI 뉴스","description":"Cursor Research는 Composer 모델인 Mixture-of-Kittens(MoK) 뒤에 있는 MoE 훈련 메가커널을 오픈소스로 공개했으며, 모든 통신과 계산을 하나의 결정론적 커널로 융합하여 가장 강력한 공개 기준선보다 최대 2.37배 빠른 속도를 달성하고, 이미 수만 개의 GPU에서 Composer 훈련을...","url":"https://www.aioga.com/ko/news/cmsf13bs51dcsro2expu0v1pt/","contentTranslated":true,"sourceHash":"c9ff0d876aa8ba41","translatedAt":"2026-08-04T20:25:32.846Z"},"es":{"title":"Cursor Open Source Mixture-of-Kittens (MoK): Megakernel de entrenamiento MoE determinista para racks GB300 NVL72","summary":"Cursor Research ha abierto el megakernel de entrenamiento MoE detrás de su modelo Composer—Mixture-of-Kittens (MoK), fusionando toda la comunicación y computación en un único núcleo determinista, alcanzando velocidades hasta 2,37 veces superiores a la base pública más fuerte, y ya soporta el entrenamiento de Composer en decenas de miles de GPUs.","category":"Productos","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"Cursor Open Source Mixture-of-Kittens (MoK): Megakernel de entrenamiento MoE determinista para racks GB300 NVL72 - Aioga Noticias de IA","description":"Cursor Research ha abierto el megakernel de entrenamiento MoE detrás de su modelo Composer—Mixture-of-Kittens (MoK), fusionando toda la comunicación y computación en un único núcle...","url":"https://www.aioga.com/es/news/cmsf13bs51dcsro2expu0v1pt/","contentTranslated":true,"sourceHash":"c9ff0d876aa8ba41","translatedAt":"2026-08-04T20:25:32.904Z"},"fr":{"title":"Cursor Open Source Mixture-of-Kittens (MoK) : Méganoyau d’entraînement MoE déterministe pour racks NVL72 GB300","summary":"Cursor Research a rendu open source le méganoyau d’entraînement MoE derrière son modèle Composer — Mixture-of-Kittens (MoK), fusionnant toute la communication et le calcul en un seul noyau déterministe, atteignant des vitesses jusqu’à 2,37 fois supérieures à la référence publique la plus puissante, et prend déjà en charge la formation Composer sur des dizaines de milliers de GPU.","category":"Produits","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"Cursor Open Source Mixture-of-Kittens (MoK) : Méganoyau d’entraînement MoE déterministe pour racks NVL72 GB300 - Aioga Actualités IA","description":"Cursor Research a rendu open source le méganoyau d’entraînement MoE derrière son modèle Composer — Mixture-of-Kittens (MoK), fusionnant toute la communication et le calcul en un se...","url":"https://www.aioga.com/fr/news/cmsf13bs51dcsro2expu0v1pt/","contentTranslated":true,"sourceHash":"c9ff0d876aa8ba41","translatedAt":"2026-08-04T20:25:40.155Z"},"de":{"title":"Cursor Open Source Mixture-of-Kittens (MoK): Deterministischer MoE-Trainings-Megakernel für GB300 NVL72-Racks","summary":"Cursor Research hat den MoE-Trainings-Megakernel hinter seinem Composer-Modell – Mixture-of-Kittens (MoK) – als Open Source veröffentlicht, indem alle Kommunikation und Berechnungen zu einem einzigen deterministischen Kernel zusammengeführt werden, Geschwindigkeiten erreicht, die bis zu 2,37-mal schneller sind als die stärkste öffentliche Basislinie, und bereits das Composer-Training auf Zehntausenden von GPUs unterstützt.","category":"产品更新","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"Cursor Open Source Mixture-of-Kittens (MoK): Deterministischer MoE-Trainings-Megakernel für GB300 NVL72-Racks - Aioga KI-News","description":"Cursor Research hat den MoE-Trainings-Megakernel hinter seinem Composer-Modell – Mixture-of-Kittens (MoK) – als Open Source veröffentlicht, indem alle Kommunikation und Berechnunge...","url":"https://www.aioga.com/de/news/cmsf13bs51dcsro2expu0v1pt/","contentTranslated":true,"sourceHash":"c9ff0d876aa8ba41","translatedAt":"2026-08-04T20:25:41.622Z"},"pt-BR":{"title":"Cursor Open Source Mixture-of-Kittens (MoK): Megakernel de treinamento determinístico de MoE para racks GB300 NVL72","summary":"A Cursor Research disponibilizou o megakernel de treinamento MoE por trás de seu modelo Composer — Mixture-of-Kittens (MoK), fundindo toda comunicação e computação em um único kernel determinístico, alcançando velocidades até 2,37 vezes superiores à base pública mais forte, e já suporta o treinamento Composer em dezenas de milhares de GPUs.","category":"产品更新","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"Cursor Open Source Mixture-of-Kittens (MoK): Megakernel de treinamento determinístico de MoE para racks GB300 NVL72 - Aioga Notícias de IA","description":"A Cursor Research disponibilizou o megakernel de treinamento MoE por trás de seu modelo Composer — Mixture-of-Kittens (MoK), fundindo toda comunicação e computação em um único kern...","url":"https://www.aioga.com/pt-BR/news/cmsf13bs51dcsro2expu0v1pt/","contentTranslated":true,"sourceHash":"c9ff0d876aa8ba41","translatedAt":"2026-08-04T20:25:49.111Z"},"ru":{"title":"Cursor Open Source Mix-of-Kittens (MoK): Детерминированный обучающий мегаядро MoE для стоек GB300 NVL72","summary":"Cursor Research открыла обучающее мегаядро MoE, основанное на своей модели Composer — Mix-of-Kittens (MoK), объединяя всю коммуникацию и вычисления в единое детерминированное ядро, достигая скоростей до 2,37 раза выше самой сильной публичной базы, и уже поддерживает обучение Composer на десятках тысяч GPU.","category":"产品更新","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"Cursor Open Source Mix-of-Kittens (MoK): Детерминированный обучающий мегаядро MoE для стоек GB300 NVL72 - Aioga Новости ИИ","description":"Cursor Research открыла обучающее мегаядро MoE, основанное на своей модели Composer — Mix-of-Kittens (MoK), объединяя всю коммуникацию и вычисления в единое детерминированное ядро,...","url":"https://www.aioga.com/ru/news/cmsf13bs51dcsro2expu0v1pt/","contentTranslated":true,"sourceHash":"c9ff0d876aa8ba41","translatedAt":"2026-08-04T20:25:50.346Z"},"ar":{"title":"مزيج القطط مفتوحة المصدر من المؤشر (MoK): نواة تدريب MoE الحتمية لرفوف GB300 NVL72","summary":"قامت شركة كورسور للأبحاث بفتح المصدر لنواة التدريب الضخمة لوزارة السحر خلف نموذج Composer الخاص بها—مزيج القطط (MoK)، حيث دمج كل الاتصالات والحوسبة في نواة حتمية واحدة، محققة سرعات تصل إلى 2.37 مرة أسرع من أقوى خط أساس عام، وتدعم بالفعل تدريب المؤلفين على عشرات الآلاف من وحدات معالجة الرسوميات.","category":"产品更新","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"مزيج القطط مفتوحة المصدر من المؤشر (MoK): نواة تدريب MoE الحتمية لرفوف GB300 NVL72 - Aioga أخبار الذكاء الاصطناعي","description":"قامت شركة كورسور للأبحاث بفتح المصدر لنواة التدريب الضخمة لوزارة السحر خلف نموذج Composer الخاص بها—مزيج القطط (MoK)، حيث دمج كل الاتصالات والحوسبة في نواة حتمية واحدة، محققة سرعات...","url":"https://www.aioga.com/ar/news/cmsf13bs51dcsro2expu0v1pt/","contentTranslated":true,"sourceHash":"c9ff0d876aa8ba41","translatedAt":"2026-08-04T20:25:59.068Z"},"hi":{"title":"कर्सर ओपन सोर्स मिक्सचर-ऑफ-बिल्ली के बच्चे (MoK): GB300 NVL72 रैक के लिए नियतात्मक MoE प्रशिक्षण मेगाकर्नेल","summary":"कर्सर रिसर्च ने अपने कंपोजर मॉडल-मिक्सचर-ऑफ-बिल्ली के बच्चे (एमओके) के पीछे एमओई प्रशिक्षण मेगाकर्नेल को ओपन-सोर्स किया है, जो सभी संचार और गणना को एक ही नियतात्मक कर्नेल में फ्यूज करता है, सबसे मजबूत सार्वजनिक बेसलाइन की तुलना में 2.37 गुना तेज गति प्राप्त करता है, और पहले से ही हजारों जीपीयू पर संगीतकार प्रशिक्षण का समर्थन करता है।","category":"产品更新","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"कर्सर ओपन सोर्स मिक्सचर-ऑफ-बिल्ली के बच्चे (MoK): GB300 NVL72 रैक के लिए नियतात्मक MoE प्रशिक्षण मेगाकर्नेल - Aioga AI समाचार","description":"कर्सर रिसर्च ने अपने कंपोजर मॉडल-मिक्सचर-ऑफ-बिल्ली के बच्चे (एमओके) के पीछे एमओई प्रशिक्षण मेगाकर्नेल को ओपन-सोर्स किया है, जो सभी संचार और गणना को एक ही नियतात्मक कर्नेल में फ्यूज...","url":"https://www.aioga.com/hi/news/cmsf13bs51dcsro2expu0v1pt/","contentTranslated":true,"sourceHash":"c9ff0d876aa8ba41","translatedAt":"2026-08-04T20:25:58.991Z"},"it":{"title":"Cursor Open Source Mixture-of-Kittens (MoK): Megakernel di addestramento MoE deterministico per rack GB300 NVL72","summary":"Cursor Research ha reso open source il megakernel di addestramento MoE dietro il suo modello Composer—Mixture-of-Kittens (MoK), fondendo tutta la comunicazione e il calcolo in un unico kernel deterministico, raggiungendo velocità fino a 2,37 volte superiori alla base pubblica più potente, e già supporta l'addestramento Composer su decine di migliaia di GPU.","category":"产品更新","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"Cursor Open Source Mixture-of-Kittens (MoK): Megakernel di addestramento MoE deterministico per rack GB300 NVL72 - Aioga Notizie IA","description":"Cursor Research ha reso open source il megakernel di addestramento MoE dietro il suo modello Composer—Mixture-of-Kittens (MoK), fondendo tutta la comunicazione e il calcolo in un u...","url":"https://www.aioga.com/it/news/cmsf13bs51dcsro2expu0v1pt/","contentTranslated":true,"sourceHash":"c9ff0d876aa8ba41","translatedAt":"2026-08-04T20:26:08.250Z"},"nl":{"title":"Cursor Open Source Mixture-of-Kittens (MoK): Deterministische MoE-training Megakernel voor GB300 NVL72-racks","summary":"Cursor Research heeft de MoE-trainingsmegakernel achter zijn Composer-model—Mixture-of-Kittens (MoK) open source gemaakt, waarbij alle communicatie en berekening worden samengevoegd tot één deterministische kernel, snelheden tot 2,37 keer sneller dan de sterkste publieke basislijn, en Composer-training al ondersteunt op tienduizenden GPU's.","category":"产品更新","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"Cursor Open Source Mixture-of-Kittens (MoK): Deterministische MoE-training Megakernel voor GB300 NVL72-racks - Aioga AI-nieuws","description":"Cursor Research heeft de MoE-trainingsmegakernel achter zijn Composer-model—Mixture-of-Kittens (MoK) open source gemaakt, waarbij alle communicatie en berekening worden samengevoeg...","url":"https://www.aioga.com/nl/news/cmsf13bs51dcsro2expu0v1pt/","contentTranslated":true,"sourceHash":"c9ff0d876aa8ba41","translatedAt":"2026-08-04T20:26:07.656Z"},"tr":{"title":"Cursor Open Source Mix-of-Kittens (MoK): GB300 NVL72 racks için deterministik MoE eğitim megakernel","summary":"Cursor Research, MoE eğitim megaçekirdeğini Composer modeli Mix-of-Kittens (MoK) arkasında açık kaynaklı olarak kullandı; tüm iletişim ve hesaplamayı tek bir deterministik çekirdekte birleştirerek, en güçlü halka açık temelden 2,37 kat daha hızlı hızlar elde etti ve zaten on binlerce GPU'da Composer eğitimini destekliyor.","category":"产品更新","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"Cursor Open Source Mix-of-Kittens (MoK): GB300 NVL72 racks için deterministik MoE eğitim megakernel - Aioga AI Haberleri","description":"Cursor Research, MoE eğitim megaçekirdeğini Composer modeli Mix-of-Kittens (MoK) arkasında açık kaynaklı olarak kullandı; tüm iletişim ve hesaplamayı tek bir deterministik çekirdek...","url":"https://www.aioga.com/tr/news/cmsf13bs51dcsro2expu0v1pt/","contentTranslated":true,"sourceHash":"c9ff0d876aa8ba41","translatedAt":"2026-08-04T20:26:16.938Z"},"vi":{"title":"Cursor Open Source Mixture-of-Kittens (MoK): Megakernel huấn luyện MoE xác định cho giá đỡ GB300 NVL72","summary":"Cursor Research đã mở mã nguồn megakernel đào tạo MoE phía sau mô hình Composer của mình—Mixture-of-Kittens (MoK), hợp nhất tất cả giao tiếp và tính toán thành một kernel xác định duy nhất, đạt tốc độ nhanh hơn tới 2,37 lần so với cơ sở công khai mạnh nhất, và đã hỗ trợ huấn luyện Composer trên hàng chục nghìn GPU.","category":"产品更新","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"Cursor Open Source Mixture-of-Kittens (MoK): Megakernel huấn luyện MoE xác định cho giá đỡ GB300 NVL72 - Tin tức AI Aioga","description":"Cursor Research đã mở mã nguồn megakernel đào tạo MoE phía sau mô hình Composer của mình—Mixture-of-Kittens (MoK), hợp nhất tất cả giao tiếp và tính toán thành một kernel xác định...","url":"https://www.aioga.com/vi/news/cmsf13bs51dcsro2expu0v1pt/","contentTranslated":true,"sourceHash":"c9ff0d876aa8ba41","translatedAt":"2026-08-04T20:26:16.900Z"},"id":{"title":"Cursor Open Source Mixture-of-Kittens (MoK): Megakernel pelatihan MoE deterministik untuk rak GB300 NVL72","summary":"Cursor Research telah melakukan open-source pada megakernel pelatihan MoE di balik model Composer—Mixture-of-Kittens (MoK), menggabungkan semua komunikasi dan komputasi menjadi satu kernel deterministik, mencapai kecepatan hingga 2,37 kali lebih cepat dari baseline publik terkuat, dan sudah mendukung pelatihan Composer pada puluhan ribu GPU.","category":"产品更新","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"Cursor Open Source Mixture-of-Kittens (MoK): Megakernel pelatihan MoE deterministik untuk rak GB300 NVL72 - Berita AI Aioga","description":"Cursor Research telah melakukan open-source pada megakernel pelatihan MoE di balik model Composer—Mixture-of-Kittens (MoK), menggabungkan semua komunikasi dan komputasi menjadi sat...","url":"https://www.aioga.com/id/news/cmsf13bs51dcsro2expu0v1pt/","contentTranslated":true,"sourceHash":"c9ff0d876aa8ba41","translatedAt":"2026-08-04T20:26:25.162Z"},"th":{"title":"เคอร์เซอร์โอเพนซอร์ส Mixture-of-Kittens (MoK): เมกะเคอร์เนลฝึก MoE แบบกําหนดสําหรับแร็ค GB300 NVL72","summary":"Cursor Research ได้เปิดซอร์สเมกะเคอร์เนลฝึกอบรม MoE ที่อยู่เบื้องหลังโมเดล Composer ของตน—Mixture-of-Kittens (MoK) โดยรวมการสื่อสารและการคํานวณทั้งหมดไว้ในเคอร์เนลที่กําหนดไว้ล่วงหน้า ทําความเร็วได้เร็วกว่ามาตรฐานสาธารณะที่แข็งแกร่งที่สุดถึง 2.37 เท่า และยังรองรับการฝึก Composer บนการ์ดจอหลายหมื่นตัวแล้ว","category":"产品更新","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"เคอร์เซอร์โอเพนซอร์ส Mixture-of-Kittens (MoK): เมกะเคอร์เนลฝึก MoE แบบกําหนดสําหรับแร็ค GB300 NVL72 - ข่าว AI Aioga","description":"Cursor Research ได้เปิดซอร์สเมกะเคอร์เนลฝึกอบรม MoE ที่อยู่เบื้องหลังโมเดล Composer ของตน—Mixture-of-Kittens (MoK) โดยรวมการสื่อสารและการคํานวณทั้งหมดไว้ในเคอร์เนลที่กําหนดไว้ล่วงห...","url":"https://www.aioga.com/th/news/cmsf13bs51dcsro2expu0v1pt/","contentTranslated":true,"sourceHash":"c9ff0d876aa8ba41","translatedAt":"2026-08-04T20:26:25.734Z"},"pl":{"title":"Cursor Open Source Mixture-of-Kittens (MoK): Deterministyczne trenowanie megajądra MoE dla szaf GB300 NVL72","summary":"Cursor Research udostępnił jako open source megajądro treningowe MoE za swoim modelem Composer — Mixture-of-Kittens (MoK), łącząc całą komunikację i obliczenia w jednym deterministycznym jądrze, osiągając prędkości do 2,37 razy szybsze niż najsilniejsza publiczna baza bazowa, a już wspiera szkolenie Composer na dziesiątkach tysięcy GPU.","category":"产品更新","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"Cursor Open Source Mixture-of-Kittens (MoK): Deterministyczne trenowanie megajądra MoE dla szaf GB300 NVL72 - Aioga Wiadomości AI","description":"Cursor Research udostępnił jako open source megajądro treningowe MoE za swoim modelem Composer — Mixture-of-Kittens (MoK), łącząc całą komunikację i obliczenia w jednym determinist...","url":"https://www.aioga.com/pl/news/cmsf13bs51dcsro2expu0v1pt/","contentTranslated":true,"sourceHash":"c9ff0d876aa8ba41","translatedAt":"2026-08-04T20:26:34.508Z"}}}}