DeepSeek To Release Next Flagship AI Model With Strong Coding Ability
The Takeaway
- DeepSeek to launch V4 AI model with strong coding ability.
- V4 model outperforms rivals in internal coding benchmarks.
- Model makes breakthroughs handling extremely long coding prompts.
Chinese AI startup DeepSeek is expected to launch its next-generation AI model that features strong coding capabilities in the coming weeks, according to two people with direct knowledge of the plan.
The new model, V4, is a successor to the V3 model DeepSeek released in December 2024. Initial tests done by DeepSeek employees based on the company’s internal benchmarks showed that it outperformed existing models, such as Anthropic’s Claude and OpenAI’s GPT series, in coding, the two people said.
DeepSeek plans to release V4 around the time of the Lunar New Year in mid-February, according to the two people, who added the timeline could still shift. DeepSeek’s last flagship model, R1, was released on Jan. 20 last year—just about a week before China was heading to the weeklong Lunar New Year celebration—ensuring it could receive abundant hype and attention.
While the V3 model put DeepSeek on the map in the global AI community, the release of R1 sent shockwaves through Silicon Valley and Wall Street and made DeepSeek a global phenomenon. R1 is an open-source “reasoning” model designed to solve complex problems by spending time to “think” through a query before providing an answer. It got enormous attention because it performed well despite the fact DeepSeek spent relatively little money on training—in contrast to leading models developed in the U.S.
In China, DeepSeek also released a chatbot, which uses both the R1 and V3 models, which quickly became popular. DeepSeek itself became a source of national pride.
DeepSeek’s breakout success intensified competitions among Chinese large-language model developers, which rushed to launch their own open-source offerings throughout 2025. From tech giants Alibaba Group, Baidu to startups MiniMax, Zhipu and Moonshot AI, their efforts to open source their models collectively elevated China’s status as the global leader of open-source AI.
The company in December released the V3.2, which outperformed OpenAI’s GPT-5 and Google’s Gemini 3.0 Pro on some benchmarks. But it hasn’t released a major successor to its models. This makes the upcoming launch of V4 particularly noteworthy.
DeepSeek didn’t respond to a request for comment.
V4 made breakthroughs in handling and processing extremely long coding prompts, a potentially significant advantage for developers working on complex software projects, the two people said. The model also made improvements in its ability to comprehend data patterns throughout every stage of the training process without degradation, they added.
AI Training requires models to repeatedly learn from massive datasets, but patterns can degrade over successive training runs. Developers armed with enormous stacks of AI chips can typically address this problem by doing more training runs.
But in the case of DeepSeek, it has to achieve better results by utilizing novel training techniques because, like other Chinese model developers, it can’t easily access the most cutting-edge chips from the U.S. But The Information reported recently that DeepSeek had got access to thousands of Nvidia’s state-of-the-art Blackwell chips, which were smuggled into China via convoluted schemes.
Users will likely notice that the V4 produces organized answers, which shows the new version has a deeper reasoning capability and will be more reliable for complex tasks, one of the people said.
Last week, the company published a research paper, co-authored by CEO Liang Wenfeng, on a new training architecture that could enable it to build much larger models without needing proportionally more chips. The developments suggest that DeepSeek has continued to innovate despite limitations imposed by U.S. export restrictions.