DeepSeek’s V4 model is expected to drop this week, according to reports from the Financial Times, Reuters, and Yahoo Finance. If the timeline holds, V4 would arrive just ahead of China’s “Two Sessions” parliamentary meetings starting March 4. That timing isn’t accidental.
According to the Financial Times, V4 will be multimodal, handling text, image, and video generation. Reuters reports that DeepSeek shared V4 with Huawei and Cambricon rather than Nvidia, a break from standard practice that signals continued hardware decoupling from U.S. chip suppliers.
V4 has not launched as of March 3. All reporting is anticipatory. Unverified benchmark leaks have circulated in community channels, but without official release or independent testing, specific performance claims can’t be confirmed.
What’s confirmed: the 14-month gap since R1 is the longest between major DeepSeek releases. R1 disrupted the frontier model conversation in January 2025 by matching Western lab performance at a fraction of the training cost. The question for V4 is whether DeepSeek can repeat that trick with multimodal capabilities while running entirely on Chinese-made chips.
The geopolitical angle is unavoidable. A Chinese AI lab releasing a competitive frontier model on domestic hardware, timed to the country’s most important political event, while Anthropic and OpenAI fight over Pentagon access. The AI competition between the U.S. and China just got a concrete data point. V4’s actual benchmarks will determine whether that data point is a headline or a footnote.