--- title: LongCat 2.0: The Beginning of the End of NVIDIA MOAT? description: LongCat 2.0 is a 1.6 trillion parameter open-weights LLM from Meituan-backed AI lab LongCat, trained entirely on Chinese AI hardware (ASIC-based AI6 superpods)... canonical: https://daily.dev/posts/longcat-2-0-the-beginning-of-the-end-of-nvidia-moat--h9ncehng4 twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: LongCat 2.0: The Beginning of the End of NVIDIA MOAT? | daily.dev og:description: LongCat 2.0 is a 1.6 trillion parameter open-weights LLM from Meituan-backed AI lab LongCat, trained entirely on Chinese AI hardware (ASIC-based AI6 superpods)... og:url: https://daily.dev/posts/longcat-2-0-the-beginning-of-the-end-of-nvidia-moat--h9ncehng4 og:image: https://api.daily.dev/og/posts/h9ncEhNg4.png og:image:alt: LongCat 2.0: The Beginning of the End of NVIDIA MOAT? og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # LongCat 2.0: The Beginning of the End of NVIDIA MOAT? **[bycloud](https://daily.dev/sources/bycloud)** 路 13 min read 路 47 upvotes 路 4 comments ## Summary LongCat 2.0 is a 1.6 trillion parameter open-weights LLM from Meituan-backed AI lab LongCat, trained entirely on Chinese AI hardware (ASIC-based AI6 superpods) without NVIDIA GPUs. The model matches near state-of-the-art performance and introduces three key architectural innovations: (1) LongCat Sparse Attention with streaming-aware indexing, cross-layer indexing with distillation, and hierarchical coarse-to-fine retrieval; (2) N-gram embeddings allocating 135B parameters to learn multi-token phrase representations up to 5-grams, particularly beneficial for code; and (3) use of the Muon optimizer and multi-tier on-policy distillation (MOPD) for post-training. The model supports a 1 million token context window and achieves 59.5% on SWEBench Pro, positioning it as a strong coding and agentic model. Its successful training on Chinese hardware with no loss spikes signals a meaningful step toward NVIDIA independence for frontier AI training. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### What hardware was LongCat 2.0 trained on instead of Nvidia GPUs? LongCat 2.0 was trained and is served entirely on a Chinese hardware stack using over 50,000 AS6 super pods, where AS6 refers to application-specific integrated circuits built for AI workloads. The blog release compares the immaturity of the AS6 superpod software ecosystem to Nvidia's GPU ecosystem, implying LongCat has become fully independent from Nvidia GPUs for both training and inferencing. _Teams weighing non-Nvidia AI hardware strategies can track cases like this on daily.dev._ ### What is engram embedding in LongCat 2.0 and how many parameters does it use? Engram embedding is a novel component in LongCat 2.0 that allocates around 135 billion parameters to learn representations for short sequences of up to five tokens, not just individual tokens, expanding the embedding space roughly 100 times. Attached to the embedding layer rather than as a transformer block, it helps capture phrase-level and code-pattern structure early in the network, keeping the engram share under 10% of the full model. _Developers exploring alternative sparsity techniques beyond mixture-of-experts can follow architecture breakdowns like this on daily.dev._ ### How does LongCat Sparse Attention differ from DeepSeek Sparse Attention? LongCat Sparse Attention is a variant of DeepSeek Sparse Attention that adds three hardware-efficiency improvements: streaming-aware indexing for contiguous memory reads, cross-layer indexing that shares one index across two layers (versus GLM 5.2's four layers) using cross-layer distillation, and hierarchical coarse-to-fine retrieval that searches block-level chunks before individual tokens, all aimed at supporting a 1 million token context window efficiently. _Anyone comparing long-context attention designs across frontier models can track these variants on daily.dev._ ## Community discussion Top comments from developers on daily.dev. **@baz14** 路 7 upvotes > This is the first major dent in Nvidia's monopoly. Expect to see this trend play out in other sectors too 馃槒 > > Hard to believe 1T+ params is already baseline now馃く > > Seeing these massive new open-weight models drop back-to-back changes everything. Open weights are a strict necessity if we want a truly fair comspetition 馃挭 **@harshaj99** 路 1 upvotes > Competition Is Better **@lorinreyes** 路 1 upvotes > Promising news for AI users. > > Competition between vendors and providers will lower the cost of AI usage and will accelerate its application. **@agustinbarrientos** 路 0 upvotes > I think one million tokens still needs a monorepo test ## Similar posts on daily.dev - [China鈥檚 Meituan says its new AI model was trained on domestic chips](https://daily.dev/posts/china-s-meituan-says-its-new-ai-model-was-trained-on-domestic-chips-sva3y3dc1) 路 The Next Web 路 0 upvotes 路 0 comments - [Claude, ChatGPT, and Gemini get all the hype, but the most interesting AI models are coming from elsewhere](https://daily.dev/posts/claude-chatgpt-and-gemini-get-all-the-hype-but-the-most-interesting-ai-models-are-coming-from-els-glbavr35d) 路 XDA Developers 路 0 upvotes 路 0 comments --- Tags: [#llm](https://daily.dev/tags/llm), [#nvidia](https://daily.dev/tags/nvidia), [#gpu](https://daily.dev/tags/gpu), [#mixture-of-experts](https://daily.dev/tags/mixture-of-experts) [View this post on daily.dev](https://daily.dev/posts/longcat-2-0-the-beginning-of-the-end-of-nvidia-moat--h9ncehng4) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"LongCat 2.0: The Beginning of the End of NVIDIA MOAT?","url":"https://daily.dev/posts/longcat-2-0-the-beginning-of-the-end-of-nvidia-moat--h9ncehng4","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/longcat-2-0-the-beginning-of-the-end-of-nvidia-moat--h9ncehng4"},"datePublished":"2026-08-08T22:02:48.944Z","dateModified":"2026-09-14T09:00:56.748Z","description":"LongCat 2.0 is a 1.6 trillion parameter open-weights LLM from Meituan-backed AI lab LongCat, trained entirely on Chinese AI hardware (ASIC-based AI6 superpods)...","image":"https://i.ytimg.com/vi/GyyXaCcCxPs/sddefault.jpg","thumbnailUrl":"https://i.ytimg.com/vi/GyyXaCcCxPs/sddefault.jpg","isAccessibleForFree":true,"articleSection":"bycloud","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"bycloud","logo":"https://media.daily.dev/image/upload/s--UUgAibgb--/f_auto/v1712315122/logos/bycloud","url":"https://daily.dev/sources/bycloud"},"commentCount":4,"discussionUrl":"https://daily.dev/posts/longcat-2-0-the-beginning-of-the-end-of-nvidia-moat--h9ncehng4","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":47},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":4}],"keywords":"llm,nvidia,gpu,mixture-of-experts","timeRequired":"PT13M","video":{"@type":"VideoObject","name":"LongCat 2.0: The Beginning of the End of NVIDIA MOAT?","description":"LongCat 2.0 is a 1.6 trillion parameter open-weights LLM from Meituan-backed AI lab LongCat, trained entirely on Chinese AI hardware (ASIC-based AI6 superpods)...","thumbnailUrl":"https://i.ytimg.com/vi/GyyXaCcCxPs/sddefault.jpg","uploadDate":"2026-08-08T22:02:48.944Z","duration":"PT13M","url":"https://api.daily.dev/r/h9ncEhNg4","embedUrl":"https://www.youtube.com/embed/GyyXaCcCxPs"}} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"bycloud","item":"https://daily.dev/sources/bycloud"},{"@type":"ListItem","position":3,"name":"LongCat 2.0: The Beginning of the End of NVIDIA MOAT?"}]} {"@context":"https://schema.org","@type":"WebPage","@id":"https://daily.dev/posts/longcat-2-0-the-beginning-of-the-end-of-nvidia-moat--h9ncehng4","comment":[{"@type":"Comment","text":"This is the first major dent in Nvidia鈥檚 monopoly. Expect to see this trend play out in other sectors too 馃槒\nHard to believe 1T+ params is already baseline now馃く\nSeeing these massive new open-weight models drop back-to-back changes everything. Open weights are a strict necessity if we want a truly fair comspetition 馃挭","datePublished":"2026-08-09T07:18:47.936Z","url":"https://daily.dev/posts/h9ncEhNg4#c-YgCGr86Uz","author":{"@type":"Person","name":"Baz","url":"https://daily.dev/baz14","image":"https://lh3.googleusercontent.com/a/ACg8ocKd2iFU8YS4sV-rBJptuYSF8EHeQnm23N_dcXE1uON7_GMIlZg=s96-c"},"interactionStatistic":{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":7}},{"@type":"Comment","text":"Competition Is Better","datePublished":"2026-08-09T16:42:32.293Z","url":"https://daily.dev/posts/h9ncEhNg4#c-g2mxji8KO","author":{"@type":"Person","name":"Harsha Lakshan Jayathissa","url":"https://daily.dev/harshaj99","image":"https://avatars.githubusercontent.com/u/112598753?v=4"},"interactionStatistic":{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1}},{"@type":"Comment","text":"Promising news for AI users.\nCompetition between vendors and providers will lower the cost of AI usage and will accelerate its application.","datePublished":"2026-08-10T04:45:16.976Z","url":"https://daily.dev/posts/h9ncEhNg4#c-idAs1ltpi","author":{"@type":"Person","name":"Lorin Reyes","url":"https://daily.dev/lorinreyes","image":"https://media.daily.dev/image/upload/s--iUqTXLqI--/f_auto/v1785636774/avatars/avatar_sjwvePDvHVHLi0ZolZR6k?_a=BAMAMicg0"},"interactionStatistic":{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1}},{"@type":"Comment","text":"I think one million tokens still needs a monorepo test","datePublished":"2026-08-21T07:57:08.124Z","url":"https://daily.dev/posts/h9ncEhNg4#c-Wr2oZbeWw","author":{"@type":"Person","name":"Agustin Barrientos","url":"https://daily.dev/agustinbarrientos","image":"https://media.daily.dev/image/upload/s--5ayxQnqn--/f_auto/v1788281802/avatars/avatar_wQYYVe5Tbj0NJ7C7qPoa8?_a=BAMAMicg0"}}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/longcat-2-0-the-beginning-of-the-end-of-nvidia-moat--h9ncehng4#faq","mainEntity":[{"@type":"Question","name":"What hardware was LongCat 2.0 trained on instead of Nvidia GPUs?","acceptedAnswer":{"@type":"Answer","text":"LongCat 2.0 was trained and is served entirely on a Chinese hardware stack using over 50,000 AS6 super pods, where AS6 refers to application-specific integrated circuits built for AI workloads. The blog release compares the immaturity of the AS6 superpod software ecosystem to Nvidia's GPU ecosystem, implying LongCat has become fully independent from Nvidia GPUs for both training and inferencing. Teams weighing non-Nvidia AI hardware strategies can track cases like this on daily.dev."}},{"@type":"Question","name":"What is engram embedding in LongCat 2.0 and how many parameters does it use?","acceptedAnswer":{"@type":"Answer","text":"Engram embedding is a novel component in LongCat 2.0 that allocates around 135 billion parameters to learn representations for short sequences of up to five tokens, not just individual tokens, expanding the embedding space roughly 100 times. Attached to the embedding layer rather than as a transformer block, it helps capture phrase-level and code-pattern structure early in the network, keeping the engram share under 10% of the full model. Developers exploring alternative sparsity techniques beyond mixture-of-experts can follow architecture breakdowns like this on daily.dev."}},{"@type":"Question","name":"How does LongCat Sparse Attention differ from DeepSeek Sparse Attention?","acceptedAnswer":{"@type":"Answer","text":"LongCat Sparse Attention is a variant of DeepSeek Sparse Attention that adds three hardware-efficiency improvements: streaming-aware indexing for contiguous memory reads, cross-layer indexing that shares one index across two layers (versus GLM 5.2's four layers) using cross-layer distillation, and hierarchical coarse-to-fine retrieval that searches block-level chunks before individual tokens, all aimed at supporting a 1 million token context window efficiently. Anyone comparing long-context attention designs across frontier models can track these variants on daily.dev."}}]} ```