Report: DeepSeek's Next-Generation Large Model V5 Plans to Use Ascend Chips for Pretraining

Domestic chips train domestic models! DeepSeek's next-generation large model is planned to be trained on Ascend chips. Source: Semiconductor Industry Observer. Published: 2026-11-03 15:56. DeepSeek chip design generates poster. Domestic computing power moves from inference into the deep waters of training.

Recently, reports emerged that DeepSeek's next-generation large model V5 will be pre-trained on Ascend chips. This news pushes the role of domestic computing power in the large-model industry chain from the inference side to the more substantial training side, and is seen as a key attempt to test whether the domestic chip ecosystem can support the full-process development of top-tier large models. Previously, Huawei's Ascend chips were mainly used for inference tasks of DeepSeek large models.

Based on publicly available information, the earliest large-scale deployment of Ascend carrying DeepSeek models occurred around the Spring Festival of 2025. Fang Xingdong, executive vice dean of the College of Media and International Culture at Zhejiang University, records in the book "The Rise of Ascend" that on January 28, 2025, Huawei's senior management issued a decision for Huawei Cloud to quickly support DeepSeek. From New Year's Eve to the second day of the Lunar New Year, Huawei and DeepSeek completed the deployment and verification of more than 3,000 Ascend chips, successfully deploying the DeepSeek-V3/R1 large models to Huawei Cloud's Gui'an Data Center. The book also mentions that DeepSeek's explosive growth greatly excited Ren Zhengfei, and throughout the Spring Festival he directed Huawei Cloud and related product teams in responding to the traffic tsunami. In January of the same year, SiliconFlow and Huawei Cloud launched DeepSeek-V3 and R1 inference services based on Ascend.

Cooperation at this stage still focused on inference deployment, with the two sides connecting Ascend computing power to DeepSeek's large-scale services as quickly as possible. Subsequent cooperation gradually extended to deeper levels. After DeepSeek-V4 was released in April 2026, Huawei said that through the two sides' "chip-model technology collaboration," the full series of Ascend supernode products could fully support the DeepSeek V4 series models. DeepSeek, meanwhile, noted on its official API page that, limited by high-end computing power, the service throughput of the V4-Pro model remains limited at present, and it expects that after Ascend 950 supernodes go on sale in batches in the second half of the year, the Pro price will be significantly reduced.

The real turning point came with systematic alignment at the software stack level. On September 30, DeepSeek announced the official open-sourcing of infrastructure components for Huawei's Ascend computing power platform, covering the TileLang high-level language compilation tool, high-performance computing libraries (DeepGEMM, FlashMLA, TileKernel, DeepSelect, etc.) and the distributed communication library (DeepEP), corresponding one-to-one with the previously open-sourced components for the Nvidia platform. Huawei, for its part, provided the jointly defined Ascend supernode SuperPoD Flex and UBL128 networking solutions, and opened the Ascend C programming interface and underlying instruction system to support TileLang's work.

The significance of this move is that DeepSeek is not only using Ascend for inference, but has also carried out systematic alignment and co-construction with Ascend at the bottom layer of the software stack, extending adaptation to the infrastructure level that supports both training and inference. This means the two sides' cooperation has upgraded from single-point deployment to deep coupling of the underlying ecosystem, laying the technical foundation for subsequently migrating training tasks to the Ascend platform.

As for the current industry landscape, the implementation of domestic computing power is mainly concentrated on the inference side, while the training side still needs improvement. The inference segment places high demands on chip engineering deployment and software stack compatibility, and domestic chips have accumulated considerable experience through large-scale services; the training side, however, imposes higher requirements on the interconnect bandwidth of computing clusters, communication library performance, and the stability of large-scale parallel training, and the cost of ecosystem migration is markedly higher.

A recent development provides preliminary confirmation for the training side. Not long ago, the AI training platform project team of Shenzhen Hetao College, together with relevant teams from Harbin Institute of Technology (Shenzhen), the Shenzhen Research Institute of Big Data, and Huawei, and relying on the Shenzhen Smart City AI computing power platform, completed full-parameter post-training of the 1.6-trillion-parameter large model DeepSeek-V4-Pro on a domestic AI computing power cluster based on Ascend 910C. It should be noted that this training still did not involve basic pre-training, but it proved that domestic computing power clusters are already capable of carrying ultra-large-scale model post-training tasks.

If DeepSeek's next-generation large model completes pre-training based on Huawei's Ascend computing power, it will for the first time answer the question of "whether China's top large models can complete pre-training tasks on domestic computing power," a question that previously had only doubts and no answers. Its significance lies not only in the computing power choice for a single model, but also in a comprehensive stress test of the domestic chip training-side ecosystem, covering multiple dimensions such as large-scale cluster stability, communication library performance, compiler toolchain maturity, and engineering tuning capability.

Looking ahead, the pre-training progress of DeepSeek V5 will largely determine the market's confidence in and valuation of domestic computing power training capabilities. If it runs successfully as scheduled, the position of domestic chips in the large-model industry chain will move from inference support to a main force in training, and the industry landscape is expected to undergo substantive change; even if the process still requires repeated debugging, this attempt itself will accumulate valuable engineering experience for the domestic computing power training-side ecosystem. The co-construction of the software stack around Ascend, large-scale deployment of supernodes, and actual verification of training tasks are all worthy of continued industry tracking.

"Ascend Rising" was written by Fang Xingdong, a Qiushi Distinguished Professor at Zhejiang University. Over six years, he conducted hundreds of hours of in-depth interviews with dozens of key figures, documenting the development history of Ascend chips and using it to examine Huawei and, more broadly, the path by which China's technology sector is breaking through with AI computing power as its entry point.

"'The sun rises and the moon ascends,' symbolizing constant striving upward and constant breakthroughs"—this is the origin of the Ascend chip's name, and it can also be said to be Huawei's spiritual core in the chip field. The book is divided into three parts, moving from the specific to the broad: from the decision to break through at a single point, to the construction of a full-stack ecosystem, and then to a systematic breakthrough from a cluster perspective. The first part covers the launch of the Da Vinci project. At the time, it was clear within Huawei that if the company still centered on traditional CPUs, it "could only follow on others' technological tracks." So the team established that "Ascend is not just a chip, but a whole-system layout with the chip at its core," and practiced the working style of "many things need no explanation—results are the answer." The second part records the difficult struggle for survival under sanctions. Relying on others' code and compilers may seem to yield quick results, but when the technology dividend peaks, the "next jump" of the technology stack will inevitably come to a halt. Facing resistance, Xu Zhijun set the tone: "Advantages are made, not innate." Huawei must build its own self-developed system, pinning its hopes on developers gathering to form an ecosystem synergy in which "many hands make the fire burn high." The third part turns to ecosystem reconstruction. During large-scale migration, Huawei realized that "the software ecosystem is the chip's most solid moat," so instead of fighting head-on over single-card process technology, it "chose to break through from the cluster perspective."

English text is being prepared.

This article is reposted from media reports or online platforms and represents the author's personal stance or views. Our reposting is for sharing only and does not mean we endorse or agree with it. If the source attribution is incorrect or your lawful rights and interests are infringed, please contact customer service promptly, and as a neutral platform service provider we will promptly correct, delete, or handle it in accordance with the law. Comments: No user comments yet. Latest News More Related Recommendations

The more agents there are, the better Intel looks 2026-11-03 16:35

Global smartphone shipments plunge 19%, revenue rises 3% against the trend 2026-11-03 16:28

Apple's robot is not on a humanoid path 2026-11-03 16:24