DeepSeek V4 has become one of the most talked-about unreleased AI models in China.
The reason is simple: the model has been widely expected for months, several Chinese media outlets have reported possible launch windows, and the market is now treating V4 as DeepSeek’s next major step after its earlier V-series and reasoning releases. But as of now, DeepSeek has not publicly released an official V4 product page, technical report, or launch announcement through its main site or API documentation. DeepSeek’s current official API documentation still points developers to DeepSeek-V3.2 for the API-facing deepseek-chat and deepseek-reasoner models, not V4.
What is confirmed
The clearest confirmed point is that DeepSeek V4 is not publicly launched yet, at least not through DeepSeek’s official API docs or public model documentation. DeepSeek’s Chinese API documentation currently says deepseek-chat and deepseek-reasoner correspond to DeepSeek-V3.2, and its update log shows ongoing V3.1 and V3.2-era updates rather than a public V4 release.
Another confirmed point is that DeepSeek has had a period of unusually high market attention around the missing V4 release. Chinese reporting from late March and mid-April repeatedly describes V4 as delayed or still pending, while official service incidents at the same time increased speculation that the company might be preparing a major model transition. The speculation was amplified after several service disruptions hit the web, app, and API products in late March.
What Chinese media are reporting
Several Chinese outlets have reported that DeepSeek V4 has slipped through multiple expected release windows. A 36Kr article published this week says V4 has been delayed several times and that DeepSeek’s overall large-version release rhythm has slowed over the past 15 months. Earlier Chinese coverage in late March also described an expected launch path that moved from around Lunar New Year, then into early March, and later into April.
Chinese media have also reported a possible late-April 2026 target. One recent report says founder Liang Wenfeng indicated internally that DeepSeek V4 was planned for late April, although the article also notes DeepSeek did not officially respond. That means this should be treated as a reported internal timeline, not a confirmed public launch date. There are also repeated media claims that V4 is meant to be a much larger architectural step than a routine model refresh. Chinese reporting has described V4 as focusing on stronger coding, long-context handling, and possibly deeper architectural changes. Some articles go much further, mentioning ideas such as million-token context, native multimodality, long-term memory, or heavy adaptation to domestic Chinese chips. But these points come from media reports and community leaks, not from DeepSeek’s official documentation, so they should be treated as unverified claims for now.
What we should be careful not to overstate
Right now, there is no official public technical report for DeepSeek V4 in the sources I found. That means key claims about model size, architecture, benchmark scores, multimodal support, memory systems, or exact launch date remain unconfirmed. Some Chinese articles and reposts cite dramatic performance or configuration details, but at least some of those reports are clearly second-hand, rumor-based, or built from community speculation rather than official release material. One example is a January DoNews item that even carries a disclaimer saying the content was generated by an AI model and is for reference only.
So the safest reading today is: V4 is highly anticipated, likely real, likely delayed, but not yet formally documented in public by DeepSeek.
Why V4 matters
Even without an official launch, V4 matters because DeepSeek has become one of the most closely watched Chinese AI companies. Expectations for V4 are not just about one more model release. They are about whether DeepSeek can keep pace with a faster industry cycle in which frontier labs are updating models, agents, APIs, and products much more frequently. 36Kr frames the issue directly: this is no longer just a benchmark race, but an industrial competition around compute, data pipelines, post-training systems, evaluation, product loops, and team scale.
That is also why the market is reading so much into every V4 delay. If V4 lands strong, it could re-establish DeepSeek as a pace-setting Chinese model company. If it slips again, the bigger question becomes whether DeepSeek is constrained less by ideas and more by the hard realities of compute, engineering capacity, and release discipline. That broader tension is a recurring theme across recent Chinese coverage.
Bottom line
What we know about DeepSeek V4 is still surprisingly limited.
What is confirmed is that DeepSeek’s public API stack is still on V3.2-era models, and that V4 has not yet appeared as an officially documented release. What is widely reported in Chinese media is that V4 has missed several expected windows, may be targeting a late-April launch, and is expected to be a more ambitious architectural upgrade rather than a small iteration. What is not confirmed is most of the exciting part: its final specs, benchmark results, multimodal capabilities, memory design, or exact release date.
For now, the best way to think about DeepSeek V4 is not as a finished product, but as a high-expectation model sitting in the gap between official silence and aggressive market rumor.