DeepSeek Unveils V4 AI Model Optimized for Huawei Ascend Chips

DeepSeek, a Hangzhou AI startup, rose to prominence in early 2025 with its R1 reasoning model, matching leading U.S. systems at lower cost. The firm originally relied on Nvidia GPUs, but U.S. export controls from 2022 spurred a shift toward domestic chips such as Huawei’s Ascend.

China’s DeepSeek Says New V4 AI Model Can Run on Huawei Chips

On 24 April 2026 DeepSeek previewed V4 in two versions. V4-Pro has 1.6 trillion total parameters (49 billion active) trained on 33 trillion tokens; V4-Flash has 284 billion total parameters (13 billion active) with 32 trillion tokens. Both support a one-million-token context window. Pricing is CNY 1 per million input tokens and CNY 12 per million output for Pro, and CNY 0.2 input / CNY 2 output for Flash. Hybrid attention reduces compute to 27 % and memory to 10 % of its predecessor for long-context tasks. Huawei confirmed support on its Ascend 950 supernode clusters, with its chips used in part of V4-Flash training.

China’s DeepSeek Unveils New AI Model Just Hours After OpenAI’s GPT-5.5 Release

DeepSeek announced on WeChat that the new V4 model will include a one-million-token context window as a standard feature for all official services. Huawei said its Ascend SuperPoD line now supports the V4 series. The White House, through adviser Michael Kratsios, claimed that foreign entities, primarily in China, are conducting industrial-scale distillation campaigns to steal American AI. Chinese foreign-ministry spokesman Guo Jiakun dismissed the allegation as baseless. U.S. officials and Anthropic have suggested DeepSeek may have used prohibited Nvidia chips and Claude outputs, which the company has not addressed.

Taipei Times

Sources differ on whether Nvidia hardware was used in V4’s training; DeepSeek disclosed only that Huawei chips contributed to V4-Flash, while earlier reports note historic reliance on Nvidia. Performance claims also vary: DeepSeek says V4-Pro trails only Google’s Gemini-3.1 Pro in world-knowledge benchmarks, whereas Reuters reports it still lags behind OpenAI’s GPT-5.4 on some tasks. Training-cost details remain undisclosed.

ReutersDeepSeek Drops V4 PreviewChina’s DeepSeek Says New V4 AI Model Can Run on Huawei ChipsTaipei Times

The launch demonstrates progress toward Chinese AI-infrastructure self-sufficiency, potentially lowering cost barriers for domestic developers and expanding long-context capabilities beyond research labs. It also sharpens the ongoing U.S.–China technology rivalry.

China’s DeepSeek Says New V4 AI Model Can Run on Huawei Chips

• “This is a big deal for China’s AI industry,” — He Hui, director, Omdia

China’s DeepSeek Says New V4 AI Model Can Run on Huawei Chips

• “From now on, a one-million-token context window will be a standard feature for all official services of DeepSeek,” — DeepSeek spokesperson, WeChat post

China’s DeepSeek Unveils New AI Model Just Hours After OpenAI’s GPT-5.5 Release

• “The US has evidence that foreign entities, primarily in China, are running industrial-scale distillation campaigns to steal American AI,” — Michael Kratsios, US science and technology adviser

China’s DeepSeek Says New V4 AI Model Can Run on Huawei Chips

• “The US claims are entirely baseless,” — Guo Jiakun, Chinese foreign-ministry spokesman

China’s DeepSeek Says New V4 AI Model Can Run on Huawei Chips

Huawei plans to scale Ascend 950 supernodes in the second half of 2026, which DeepSeek expects will reduce Pro pricing. A U.S.–China summit in Beijing is slated for next month. DeepSeek will retire its older models by July 2026 and continue refining V4.

Taipei Times