DeepSeek Prepares to Compete with AI Giants with V4 Model
mHC Technology Could Outperform GPT and Claude
Chinese company DeepSeek has announced the imminent release of its new V4 language model, which, according to insider information, is capable of surpassing OpenAI and Anthropic solutions in programming tasks.
As reported by The Information citing sources within the company, the official launch is scheduled for mid-February 2026—approximately February 17, during the Chinese New Year celebrations.
The key feature of V4 is its outstanding capabilities in code processing, especially when working with very long prompts. According to DeepSeek's internal test results, the new model outperforms Anthropic's Claude and OpenAI's GPT series in programming tasks.
To become a leader in this field, V4 must surpass the current record of Claude Opus 4.5 in the SWE-bench Verified benchmark, where the Anthropic model demonstrates a result of 80.9%.
According to sources, the breakthrough may be enabled by the Manifold-Constrained Hyper-Connections (mHC) technology, described in DeepSeek's recent research paper dated January 1, 2026. The method addresses a fundamental problem in scaling large language models: how to expand the architecture without losing training stability.
Unlike the previous Hyper-Connections (HC) concept from ByteDance, mHC adds mathematical constraints on the redistribution of information between neural network layers. This allows maintaining training stability even with a significant increase in the number of layers, preventing gradient explosions, and efficiently scaling the model without a proportional increase in computational resource requirements.
Wei Sun, Chief AI Analyst at Counterpoint Research, called mHC a "stunning breakthrough" that allows DeepSeek to "bypass computational power limitations and make leaps in intelligence" even under U.S. export restrictions on advanced chips.
DeepSeek has already demonstrated impressive results with previous releases. In December 2024, the V3 model scored 90.2% on the MATH-500 benchmark, significantly surpassing Claude's result of 78.3%. In January 2025, the R1 model outperformed OpenAI's o1 solution on mathematics and reasoning benchmarks. Notably, the development of R1 cost the company only $6 million—68 times cheaper than competitors.
Currently, DeepSeek has not officially confirmed information about V4's superiority over competitors, and public benchmarks have not yet been published. However, expectations around the release are high: if the claimed characteristics are confirmed, the model could become a serious player in the global AI market.