This is not a launch announcement. DeepSeek V4 has not been released publicly as of the time of writing.
What exists: an extended period of anticipation following DeepSeek’s release of R1 in January 2026, ongoing developer community discussion based on signals in DeepSeek’s published research papers and open-source repository activity, and a pattern of DeepSeek releasing models with limited advance notice. The developer community estimates that V4 is in a pre-release or near-completion phase, though DeepSeek has made no official announcement confirming a timeline.
Architecture details circulating in community discussion are sourced from DeepSeek’s published research papers, not from confirmed V4 specifications. The model lineage from V2 through R1 has established a pattern: each release improved on its predecessor’s efficiency-to-performance ratio, with the mixture-of-experts architecture allowing DeepSeek models to activate only a fraction of total parameters per inference pass. Whether V4 continues that pattern or introduces a structural departure is not known.
The policy context makes the timing significant. The Trump administration’s draft chip export framework — reported the same week — is partly premised on the competitive threat from Chinese AI development. DeepSeek’s R1 already demonstrated that strong model performance is achievable at lower compute cost than U.S. frontier labs have required. V4, whenever it arrives, will be evaluated in that frame: not just as a model release, but as a data point in the U.S.-China AI competition.
When V4 launches, full coverage will appear here. Until then, this brief stands as a status check on an anticipated release that has not yet materialized.