Amid the heated Spring Festival release season for domestic large models, DeepSeek AI’s next‑generation flagship model V4 has once again become a focal point. On 2026‑02‑11, multiple international media reported that DeepSeek V4 is expected to be officially released in mid‑February, possibly coinciding with the Lunar New Year on February 17, a development that has generated broad anticipation across the global developer community. Although DeepSeek has not officially confirmed the news, internal tests and leaked information indicate that V4 performs strongly on programming tasks and has surpassed the Claude and GPT series in internal benchmarks, particularly when handling ultra‑long code prompts. This “reportedly launching” rumor is not unfounded: as early as January 9, The Information exclusively reported that DeepSeek planned to release a code‑focused new model in February. As a representative Chinese AI startup, DeepSeek’s potential V4 release could further narrow the gap with international giants and drive a reshaping of the open‑source AI ecosystem. Based on authoritative reports, we analyze the background and technical highlights of this development.
DeepSeek’s Rise: The Open‑Source Path from V2 to V3
DeepSeek AI was founded in 2023 and is headquartered in Hangzhou. It was established by a team of former Alibaba and Tencent engineers and focuses on efficient AI model development. The company is known for “efficient training” and an “open‑source first” approach. Early products such as DeepSeek‑V2, released in 2024, had 236B parameters yet achieved state‑of‑the‑art performance at very low cost. V2 led on math and coding benchmarks and, with knowledge cutoff up to mid‑2024, has been widely used in education and enterprise applications.
In 2025, DeepSeek launched the V3 series, including V3 and V3.2, expanding parameter scale to 389B and pretraining data to 28.5T. V3 introduced sparse attention (DSA) and multi‑token prediction (MTP), significantly improving long‑sequence handling and extending the context window from 128K to 1M tokens, with a knowledge cutoff updated to May 2025. This upgrade made V3 more stable on complex tasks and avoided late‑training performance degradation. DeepSeek’s open‑source strategy has been key to its success: V3’s code was published on GitHub, attracting a global developer community and spawning derivative applications such as DeepSeek‑Coder. The company emphasizes “efficient compute,” claiming training costs at just 10% of comparable models, which helped DeepSeek stand out in resource‑constrained environments.
The “reportedly launching” of V4 stems from multiple reports. On January 25, Atlas Cloud analysis suggested V4 was expected in February with a focus on programming optimization. A Medium post further indicated V4 might debut on February 17 (Lunar New Year), a timing consistent with DeepSeek’s customary “holiday release” strategy. Reuters and Nasdaq reported that DeepSeek insiders said V4 had entered final testing. Although DeepSeek did not respond publicly, recent updates to the company’s app (such as a 1M context window and a knowledge cutoff of May 2025) have been seen as signs of V4.
Technical Expectations for V4: Engram Memory and Ultra‑Long Code Handling
According to leaks and analysis, DeepSeek V4 will inherit V3’s DSA and MTP mechanisms while introducing an innovative Engram memory module. This architecture was disclosed on January 13 and is said to efficiently retrieve information from contexts larger than 1M tokens while reducing hardware costs by 90%. Engram is described as a conditional memory technique that allows the model to stably learn data patterns during training and avoid late‑stage degradation. V4’s parameter scale is expected to reach the hundreds of billions, with expanded pretraining data and a clear emphasis on coding tasks.
High expectations surround V4’s programming ability. Internal tests reportedly show it excels at handling “ultra‑long code prompts,” capable of generating complex software architectures and fixing bugs, outperforming the Claude and GPT families. For example, on the SWE‑Bench benchmark, V4 is projected to score higher than V3’s 77.8%, approaching Claude Opus 4.5’s 80.9%. Vertu’s analysis of a leaked “MODEL1” code suggests V4 supports multi‑agent collaboration and tool invocation, suitable for real‑world development scenarios. Additionally, V4 is said to have strengthened pattern understanding in data and a more stable training process, avoiding the “late collapse” seen in some other models.
Compared with other models, V4’s open‑source nature is its biggest advantage. It is expected to offer a free API and open‑source its code on Hugging Face. Unlike contemporaries such as Qwen 3.5 and GLM‑5, V4 reportedly prioritizes cost efficiency, with training hardware requirements reduced by 90%, making it suitable for deployment by small and medium‑sized enterprises. A Substack post suggested V4’s release would spark debate about an “open‑source victory” and could drive volatility in AI‑related stocks.
Market Impact: Spring Festival Competition and the Global AI Landscape
The reported launch of DeepSeek V4 comes amid the peak of domestic AI releases for the Spring Festival. The South China Morning Post reported that DeepSeek’s updates expanded the context window to 1M tokens with a knowledge cutoff of May 2025, seen as a prelude to V4. V4 is positioned alongside Alibaba’s Qwen 3.5, ByteDance’s Doubao 2.0 and Minimax 2.5, forming a “three‑way rivalry” with these models expected to be released around the festival. Motley Fool analysis suggested V4’s programming strengths could “shake up the market again,” prompting investors to re‑evaluate AI stocks.
From a global perspective, V4’s open‑source state‑of‑the‑art status would challenge OpenAI and Anthropic. The concerted efforts of Chinese AI labs are viewed as a signal of catching up. LinkedIn posts point out that V4’s Engram technology could reshape training paradigms. Manifold Markets gave V4 a 49% probability of being released before March.
Outlook: Challenges and Opportunities Coexist
The reported launch of DeepSeek V4 signals a shift from following to potentially leading for Chinese AI, but challenges remain: open‑source efforts must balance safety, and claims of programming SOTA require independent verification. The company reportedly plans public testing on platforms such as Chatbot Arena to collect feedback.
▶️ Enjoy this article? Creating content is hard work!
Follow our YouTube channel:
AI Rosetta
Your likes, subscriptions, and shares are our greatest support! 🙏
For more AI news, visit:
https://www.airosetta.com