Huawei Founder Congratulates DeepSeek on Ascend-Based Model, Opens Access Companywide

A purported email from Huawei founder Ren Zhengfei to the founders of DeepSeek, dated April 23, 2026, has surfaced online. The email congratulates the DeepSeek team on the model trained on the Ascend computing platform and states that the model has been opened to all Huawei employees. It also discloses that Huawei and DeepSeek have established joint laboratories in Shenzhen, Dongguan, and Shanghai to continuously refine model capabilities.

On April 24, 2026, DeepSeek-V4 Preview was officially released in two versions: V4-Pro adopts a 1.6-trillion-parameter MoE architecture with 49 billion activated parameters, while V4-Flash has 284 billion total parameters with 13 billion activated. The model natively supports a 1-million-token long context and, for the first time, is deeply adapted to Huawei's Ascend chip series, completing CANN low-level operator transformations for native deployment on domestic compute power. Previously, during the 2025 Spring Festival period, Huawei Cloud's Gui'an data center urgently deployed more than 3,000 Ascend 910B chips to support DeepSeek's traffic demand.

Huawei's Ascend project was initiated in 2017 and continued investment after external restrictions in 2019. In September 2025, the Ascend 950 was released; in July 2026, the Ascend 950 supernode was unveiled at the World Artificial Intelligence Conference, offering a 1,024-card scale with total computing power of 1 EFLOPS FP8 and 2 EFLOPS FP4. By the end of 2025, CANN had completed open-sourcing, with over 12.44 million lines of open-source code and monthly active developers exceeding 3,500.

The Global Artificial Intelligence Innovation Index Report, published in July 2026, ranks the United States first with 78.44 points, followed by China with 60.49. The report notes that China's large-model training compute efficiency is on average 30.4% higher than leading US models, and token pricing for top domestic models is roughly one-quarter of comparable US products. Meanwhile, the US maintains leadership in code programming, agents, and mathematical reasoning, and its stock of intelligent computing power remains larger than China's. China has surpassed the US in model downloads and the number of derived models.

The email and release information indicate that cooperation between Huawei and DeepSeek has entered a substantive phase. The laboratories, established in Shenzhen, Dongguan, and Shanghai, will optimize models based on Ascend clusters and plan a series of training and hands-on sharing sessions. This collaboration marks deep synergy between China's computing infrastructure and large-model technology in the field of artificial intelligence.