Chinese AI startup DeepSeek has launched a preview version of its long-awaited V4 model, ending months of silence from one of China's most closely watched AI labs and arriving a year after its R1 release sparked U.S. market turmoil and concerns across Silicon Valley AI firms.
The rollout signals that DeepSeek is full steam ahead in the frontier-model race:
DeepSeek-V4 Preview is officially live & open-sourced! Welcome to the era of cost-effective 1M context length.
-
DeepSeek-V4-Pro: 1.6T total / 49B active params. Performance rivaling the world's top closed-source models.
-
DeepSeek-V4-Flash: 284B total / 13B active params. Your fast, efficient, and economical choice.
The open-source model comes in the V4 Flash and V4 Pro series, with DeepSeek saying its V4 "leads all current open models, trailing only Gemini-3.1-Pro."
In terms of reasoning, the startup said it "beats all current open models in Math/STEM/Coding, rivaling top closed-source models."
Counterpoint Research Vice President Neil Shah told CNBC that "DeepSeek's V4 preview is a serious flex."
According to Counterpoint Principal AI Analyst Wei Sun, V4's benchmark profile suggests the model could deliver "excellent agent capability at significantly lower cost."
Ivan Su, Senior Equity Analyst at Morningstar, told CNBC that V4's debut is unlikely to deliver the same market shock as R1 did a little more than a year ago, largely because Wall Street has already priced in the view that Chinese AI can be built and deployed at a lower cost.
Goldman analyst Christopher Moniz commented on the market reaction in overnight trading in China, where "GPU and domestic chip stocks rallied after DeepSeek unveiled preview versions of its latest V4 AI model."
Moniz continued, "News of DeepSeek's latest AI model created weakness in AI application names - Minimax (100 HK) -9.4% and Knowledge Atlas (2513 HK) -9.1%. Conversely, China domestic chipmakers spiked - HHS (1347 HK) +15.2% and SMIC (981 HK) +10%. Tencent (700 HK) -0.4% on mixed feedback regarding its new Hy3 model release, with locals noting it is less efficient than Minimax M2.7 launched a month ago"
One key question is the chip stack behind V4: which chips it was trained on and which hardware it runs on during inference. Huawei has already claimed that its latest AI computing cluster, powered by Ascend AI processors, can support V4, suggesting that Beijing's domestic AI hardware ecosystem is powering DeepSeek's ongoing frontier-model push.