--- title: DeepSeek open-sources six Huawei Ascend projects, led by... description: DeepSeek has open-sourced six projects enabling its GPU kernel code to run on Huawei Ascend chips instead of Nvidia hardware. The flagship project, TileLang,... canonical: https://daily.dev/posts/deepseek-open-sources-six-huawei-ascend-projects-led-by-tilelang-e1ef89jaz twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: DeepSeek open-sources six Huawei Ascend projects, led by TileLang | daily.dev og:description: DeepSeek has open-sourced six projects enabling its GPU kernel code to run on Huawei Ascend chips instead of Nvidia hardware. The flagship project, TileLang,... og:url: https://daily.dev/posts/deepseek-open-sources-six-huawei-ascend-projects-led-by-tilelang-e1ef89jaz og:image: https://api.daily.dev/og/posts/E1EF89JAz.png og:image:alt: DeepSeek open-sources six Huawei Ascend projects, led by TileLang 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. # DeepSeek open-sources six Huawei Ascend projects, led by TileLang **[Collections](https://daily.dev/sources/collections)** · 2 min read · 9 upvotes · 0 comments ## Summary DeepSeek has open-sourced six projects enabling its GPU kernel code to run on Huawei Ascend chips instead of Nvidia hardware. The flagship project, TileLang, is the language DeepSeek uses to write kernels for training its V4 models, and most operators now have Ascend-compiled versions from the same source. DeepGEMM-Ascend reports matrix multiplication at 431 BF16 TFLOPS against a 432 ceiling, the basis for the touted 99.8% peak efficiency figure. DeepEP-Ascend, which handles inter-chip data movement, shipped without a license file, unlike the rest of the MIT-licensed repos, meaning it currently cannot be legally reused. DeepSeek's founder reportedly told investors Huawei could begin delivering training chips this quarter, and one observer noted the work shows DeepSeek training language models on Ascend hardware. ## Content DeepSeek has open-sourced a set of programming tools for Huawei's Ascend AI chips. The release is a clear step toward cutting its reliance on Nvidia. ## What was released There are six projects. They cover compute and communication modules, and they mirror tools DeepSeek already published for Nvidia hardware. Huawei reportedly gave full support during development. The headline piece is TileLang, the language DeepSeek writes its own GPU kernels in. The company positions it as a simpler alternative to Nvidia's CUDA. Most of the operators used to train its V4 models are written in TileLang, and each now has an Ascend version. The same source code can compile for Huawei chips instead of Nvidia ones. ## The numbers The matrix kernels reach 99.8% of the chip's peak speed. DeepGEMM-Ascend, which handles matrix multiplication, reports 431 BF16 TFLOPS against a ceiling of 432. One catch: DeepEP-Ascend, which moves data between chips, shipped without a license file. The other projects use the MIT license. Without a license, nobody can legally reuse it yet. ## The wider picture The two companies have also built a "supernode" system from 128 Ascend 950 chips. They plan further collaboration, including deploying at least 160,000 Huawei accelerators in a data centre in Inner Mongolia. The move is framed as part of a push for an "independent, self-controlled" GPU software ecosystem, while US export controls limit Chinese access to Nvidia's most advanced chips. Liang Wenfeng told investors that Huawei could start delivering training chips this quarter. @yacineMTB noted that DeepSeek ended up doing this on Ascends for training language models, which is the part that matters most to me. Inference ports are common. Training is the harder test, and the V4 operators now having Ascend versions suggests DeepSeek is serious about it. ## Questions this post answers ### What performance does DeepSeek's DeepGEMM-Ascend achieve on Huawei Ascend chips? DeepGEMM-Ascend reports 431 BF16 TFLOPS against a theoretical ceiling of 432 TFLOPS on Huawei Ascend chips, which works out to roughly 99.8% of peak hardware speed. This is one of six projects DeepSeek open-sourced to run its GPU kernel code, originally written for Nvidia hardware, on Ascend chips instead. _Engineers evaluating Ascend as an Nvidia alternative can follow kernel benchmarks like this on daily.dev._ ### Can I use DeepSeek's DeepEP-Ascend project in my own code? Not legally yet. DeepEP-Ascend, the project handling data movement between chips, shipped without a license file, while the other five projects in the same release, including TileLang, are MIT licensed. Until a license is added, the code cannot be legally reused despite being publicly visible on the repository. _Developers tracking open-source licensing gaps in AI infra releases can follow updates like this on daily.dev._ ### What is TileLang and why did DeepSeek release an Ascend version? TileLang is the programming language DeepSeek uses to write its own GPU kernels, and most of the operators that trained its V4 models are written in it. DeepSeek released Ascend-compiled versions of these operators from the same source code, letting the same kernel code target Huawei Ascend chips instead of Nvidia GPUs, reducing reliance on Nvidia hardware. _Teams weighing hardware portability for ML training can track moves like this on daily.dev._ ## Similar posts on daily.dev - [DeepSeek V4 to run on Huawei silicon as China builds its own AI stack](https://daily.dev/posts/deepseek-v4-to-run-on-huawei-silicon-as-china-builds-its-own-ai-stack-esd3mzwej) · TechCentral · 0 upvotes · 0 comments - [DeepSeek is reportedly designing its own AI chip to sidestep US curbs](https://daily.dev/posts/deepseek-is-reportedly-designing-its-own-ai-chip-to-sidestep-us-curbs-gn1slrknd) · The Next Web · 0 upvotes · 0 comments --- Tags: [#nvidia](https://daily.dev/tags/nvidia), [#deepseek](https://daily.dev/tags/deepseek) [View this post on daily.dev](https://daily.dev/posts/deepseek-open-sources-six-huawei-ascend-projects-led-by-tilelang-e1ef89jaz) ```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":"DeepSeek open-sources six Huawei Ascend projects, led by TileLang","url":"https://daily.dev/posts/deepseek-open-sources-six-huawei-ascend-projects-led-by-tilelang-e1ef89jaz","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/deepseek-open-sources-six-huawei-ascend-projects-led-by-tilelang-e1ef89jaz"},"datePublished":"2026-10-01T05:20:54.025Z","dateModified":"2026-10-01T21:17:45.437Z","description":"DeepSeek has open-sourced six projects enabling its GPU kernel code to run on Huawei Ascend chips instead of Nvidia hardware. The flagship project, TileLang,...","image":"https://pbs.twimg.com/media/HThWIebagAAEIC7.jpg","thumbnailUrl":"https://pbs.twimg.com/media/HThWIebagAAEIC7.jpg","isAccessibleForFree":true,"articleSection":"Collections","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":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/deepseek-open-sources-six-huawei-ascend-projects-led-by-tilelang-e1ef89jaz","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":9},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"nvidia,deepseek","timeRequired":"PT2M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"DeepSeek open-sources six Huawei Ascend projects, led by TileLang"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/deepseek-open-sources-six-huawei-ascend-projects-led-by-tilelang-e1ef89jaz#faq","mainEntity":[{"@type":"Question","name":"What performance does DeepSeek's DeepGEMM-Ascend achieve on Huawei Ascend chips?","acceptedAnswer":{"@type":"Answer","text":"DeepGEMM-Ascend reports 431 BF16 TFLOPS against a theoretical ceiling of 432 TFLOPS on Huawei Ascend chips, which works out to roughly 99.8% of peak hardware speed. This is one of six projects DeepSeek open-sourced to run its GPU kernel code, originally written for Nvidia hardware, on Ascend chips instead. Engineers evaluating Ascend as an Nvidia alternative can follow kernel benchmarks like this on daily.dev."}},{"@type":"Question","name":"Can I use DeepSeek's DeepEP-Ascend project in my own code?","acceptedAnswer":{"@type":"Answer","text":"Not legally yet. DeepEP-Ascend, the project handling data movement between chips, shipped without a license file, while the other five projects in the same release, including TileLang, are MIT licensed. Until a license is added, the code cannot be legally reused despite being publicly visible on the repository. Developers tracking open-source licensing gaps in AI infra releases can follow updates like this on daily.dev."}},{"@type":"Question","name":"What is TileLang and why did DeepSeek release an Ascend version?","acceptedAnswer":{"@type":"Answer","text":"TileLang is the programming language DeepSeek uses to write its own GPU kernels, and most of the operators that trained its V4 models are written in it. DeepSeek released Ascend-compiled versions of these operators from the same source code, letting the same kernel code target Huawei Ascend chips instead of Nvidia GPUs, reducing reliance on Nvidia hardware. Teams weighing hardware portability for ML training can track moves like this on daily.dev."}}]} ```