Two days after its initial preview, DeepSeek has completed the full open-source release of its V4 model family, publishing weights for both the V4 Pro and V4 Flash variants under the permissive MIT license. The release landed on Saturday, April 26, hours after OpenAI had shipped its own GPT-5.5 on Thursday, a timing that, intentional or not, has become a signature of how the US-China AI rivalry now plays out in real time.
In a separate statement issued the same day, Huawei confirmed that its Ascend chips and related software stack are fully compatible with the DeepSeek V4 models. The announcement marks a concrete step in DeepSeek’s stated goal of reducing reliance on Nvidia hardware, a goal that has taken on new urgency as US export controls continue to restrict access to the most advanced American accelerators. DeepSeek’s V4 was built and trained on Huawei Ascend infrastructure, and Huawei’s public confirmation of compatibility signals that the Ascend ecosystem is now capable of supporting frontier-class open-source inference at scale.
V4 Pro and Flash: What the Benchmarks Show
Both V4 variants ship with a 1-million-token context window, a significant leap from the 128,000-token limit of V3. DeepSeek describes the V4 Pro Max version as delivering “superior performance” relative to OpenAI’s GPT-5.2 and Google’s Gemini 3.0-Pro on standard reasoning benchmarks, while falling “marginally” short of GPT-5.4 and Gemini 3.1-Pro. On agentic tasks — the ability to complete complex, multi-step workflows autonomously — DeepSeek says V4 Pro outperforms Anthropic’s Claude Sonnet 4.5 and approaches the level of Claude Opus 4.5 based on its own internal evaluation.
The V4 Flash variant is designed for speed and cost efficiency. DeepSeek says Flash performs on par with Pro on simple agent tasks and has reasoning capabilities that closely approach the Pro version — a positioning that mirrors the structure of competing model families from Anthropic and Google, where a smaller, cheaper model handles the bulk of production traffic while the flagship handles complex tasks.
Lian Jye Su, chief analyst at the technology research group Omdia, told AP that “based on the benchmark results, it does appear DeepSeek V4 is going to be very competitive against its US rivals.” Marina Zhang, an associate professor at the University of Technology Sydney, called the V4 rollout a “pivotal milestone for China’s AI industry,” particularly amid intensifying global competition over AI self-reliance.
Not all analysts are equally enthusiastic. Ivan Su, a senior equity analyst at Morningstar, described V4 as a “competent follow-up” but cautioned that it is “not as big a breakthrough as the rollout of R1.” Su noted that domestic competition within China has intensified significantly since R1’s release in January 2025, and that independent evaluations, rather than DeepSeek’s own benchmarks, will be needed before final conclusions can be drawn about V4’s true capabilities.
The Huawei Dimension
The Huawei compatibility announcement carries implications that extend well beyond DeepSeek’s own deployment plans. A Microsoft report published in January showed that DeepSeek adoption has been gaining ground in many developing nations, particularly in those where Huawei phones are widely used. With V4 now confirmed to run on Ascend hardware, those same markets gain access to a frontier-class open-source model that can be deployed entirely outside the Nvidia-dominated computing ecosystem.
Marina Zhang described this as a demonstration of “technical feasibility of operating outside the Nvidia-dominated computing ecosystem amid sustained technological decoupling between China and the US.” That framing is significant: it is not merely a product launch but a proof point that an alternative hardware-software stack for frontier AI is now functional and publicly accessible.
(Related: DeepSeek V4 Targets Late April Launch on Huawei Ascend Chips)
The timing relative to GPT-5.5 also deserves attention. OpenAI released GPT-5.5 on Thursday, April 24. DeepSeek’s full V4 open-source release followed within 48 hours. The pattern mirrors what happened in January 2025, when DeepSeek’s R1 release followed closely on the heels of OpenAI announcements. Whether the timing is deliberate or coincidental, it keeps DeepSeek in the same news cycle as its US rivals and reinforces the perception that the gap between Chinese and American frontier AI is narrowing in real time.
From Preview to Production
The April 26 open-source release completes a process that began with the April 24 preview. The preview gave developers and researchers an initial look at V4’s capabilities and architecture; the full release makes the weights publicly available for download, fine-tuning, and deployment. The MIT license means there are no restrictions on commercial use, which is a deliberate choice that distinguishes DeepSeek from Anthropic and OpenAI, both of which keep their frontier weights proprietary.
Anthropic has previously accused DeepSeek and two other Chinese AI laboratories of “industrial-scale campaigns” to extract capabilities from Claude using a technique called distillation, in which a less capable model is trained on the outputs of a stronger one. OpenAI made similar allegations in a letter to US lawmakers. This week, Michael Kratsios, chief science and technology adviser to President Trump, accused foreign tech companies “principally based in China” of distilling leading US AI systems and “exploiting American expertise and innovation.” China’s embassy in Washington described the allegations as “unjustified suppression of Chinese companies by the US.”
The open-source release of V4 does not resolve those allegations, but it does change the strategic landscape: any capabilities embedded in V4, however they were developed, are now available to every developer with a GPU cluster and an internet connection.
(Related: DeepSeek Releases V4 with 1.6 Trillion Parameters and Huawei’s Full Support)