DeepSeek is preparing to launch its next-generation artificial intelligence (AI) model, V4, in mid-February, according to a report in The Information, citing two people with direct knowledge of the company’s plans.
The upcoming release marks the Chinese startup’s first major model update since its R1 reasoning model in January 2025 sparked global attention and sent shockwaves through Silicon Valley.
Internal testing by DeepSeek employees suggests V4 could surpass leading competitors including Anthropic’s Claude and OpenAI’s GPT series in coding tasks, according to the report.
The new model has reportedly achieved breakthroughs in handling extremely long coding prompts, a capability that could prove valuable for developers working on complex software projects. V4 also demonstrates improved ability to comprehend data patterns throughout the training process without degradation — a common challenge in AI development that typically requires extensive computational resources to overcome.
Users can expect more organized responses from V4, indicating deeper reasoning capabilities and increased reliability for complex tasks, according to one source. This advancement is particularly notable given the constraints DeepSeek faces in accessing cutting-edge AI chips due to U.S. export restrictions.
DeepSeek did not respond to requests for comment. The timeline for the V4 release could still shift, the sources cautioned.
“From a market perspective, DeepSeek V4 is being watched as a signal of how fast coding-centric models are maturing into development engines and platforms. The question is whether these gains translate into fewer handoffs, less rework, and tighter human oversight across planning, implementation, debugging, and deployment,” said Mitch Ashley, vice president and practice lead, DevOps and AppDev, at The Futurum Group. “If DeepSeek can demonstrate that kind of practical lift, it reframes competition around software execution capability, not model size or novelty, and accelerates expectations for what AI-centric development should look like going into 2026.”
Last week, DeepSeek published a research paper co-authored by CEO Liang Wenfeng detailing a new training architecture that could enable the company to build larger models without proportionally increasing chip requirements. The developments suggest DeepSeek continues to innovate despite technological limitations.
The Hangzhou-based startup has emerged as a central figure in China’s efforts to build an independent AI ecosystem. Its December 2024 release of V3 gained international recognition, but the January launch of R1, an open-source reasoning model, transformed DeepSeek into a global phenomenon. R1 garnered widespread attention for performing competitively with leading U.S. models despite reportedly requiring far less training expenditure.
DeepSeek’s success has intensified competition among Chinese AI developers, with tech giants Alibaba and Baidu, alongside startups like MiniMax, Zhipu, and Moonshot AI, rushing to release their own open-source models throughout 2025. These collective efforts have positioned China as a global leader in open-source AI development.
The company released V3.2 in December, which outperformed OpenAI’s GPT-5 and Google’s Gemini 3.0 Pro on certain benchmarks, making the V4 launch particularly anticipated.
