DeepSeek's new model is here?
DeepSeek is advancing the grayscale testing of its new version model, which may be the ultimate grayscale edition before the official launch of V4.
On February 11th, some users received a prompt to update the DeepSeek App upon opening it. After updating the app (version 1.7.4), users can experience DeepSeek's latest model.This upgrade extends the model's context length from 128K to 1M, an almost tenfold increase; the knowledge base is updated to May 2025, and several core capabilities have been significantly improved.
Actual tests show that DeepSeek, in its Q&A, states the current version is likely not V4,but is very likely the final evolved form of the V3 series, or the ultimate grayscale edition before the official reveal of V4.
Nomura Securities released a report on February 10th saying,the DeepSeek V4 model expected to be launched in mid-February 2026 will not recreate the global panic for AI compute demand seen last year with the launch of V3. The firm believes the core value of V4 lies in driving the commercialization of AI applications through innovations in the underlying architecture, rather than disrupting the existing AI value chain.
According to evaluations, the new version's ability to handle complex tasks has matched mainstream closed-source models such as Gemini 3 Pro and K2.5. Nomura further pointed out that V4 is expected to introduce two innovative technologies, mHC and Engram, breaking the bottleneck of computing chips and memory from algorithmic and engineering levels. Initial internal tests show V4's performance in programming tasks exceeds that of Anthropic Claude and OpenAI GPT models in the same generation.
The key significance of this release lies in further compressing training and inference costs, providing a feasible path for global large language model and AI application enterprises to ease capital expenditure pressures.
Innovative architecture optimized for hardware bottlenecks
Nomura Securities' report points out that the performance of computing chips and the bottleneck of HBM memory have always been hard constraints for the domestic large model industry. The soon-to-be-released DeepSeek V4 introduces mHC (super connectivity and manifold-constrained super connectivity) and Engram architecture, which are system-level optimizations from both training and inference perspectives targeting these weaknesses.
mHC:Full name is "manifold-constrained super connectivity". Its purpose is to address the bottleneck of information flow and instability in training when Transformer models become very deep.Simply put, it makes the "conversation" between neural network layers richer and more flexible, and uses strict mathematical "guardrails" to prevent information from being amplified or damaged. Experiments show that models adopting mHC perform better in tasks such as mathematical reasoning.
Engram:A "conditional memory" module. Its design decouples "memory" from "computation".Static knowledge in the model (such as entities, fixed expressions) is specially stored in a sparse memory table, which can be placed in inexpensive DRAM. When reasoning is needed, it can be quickly retrieved. This frees up expensive GPU memory (HBM), allowing it to focus on dynamic computation.
mHC technology improves training stability and convergence efficiency, offsetting generational differences in bandwidth and computing density for domestic chips to some extent; the Engram architecture focuses on reconstructing memory scheduling mechanisms, breaking through GPU memory capacity and bandwidth limits with more efficient access strategies under constrained HBM supply. Nomura believes these two innovations together form an adaptation solution for domestic hardware ecosystems, with clear engineering implementation value.
The report further points out, the most direct business impact of V4’s release is a substantial decrease in training and inference costs. Optimization on the cost side will stimulate downstream application demand, thereby igniting a new cycle of AI infrastructure construction. In this process, Chinese AI hardware vendors are expected to benefit from dual boosts of demand expansion and investment brought forward.
Market pattern shifts from “dominance by one” to “warring factions”
Nomura's report reviewed the changes in market structure one year after the release of DeepSeek-V3/R1. By the end of 2024, DeepSeek's two models occupied more than half of the token usage by open-source models on OpenRouter.
But by the second half of 2025, as more players join, their market share has dropped significantly. The market has shifted from "dominance by one" to "warring factions". The competitive environment facing V4 is much more complex than a year ago. DeepSeek's "compute management efficiency" plus "performance upgrades" has accelerated the development of China's large language models and applications, also changing the global competitive landscape and driving more attention to open-source models.
Software companies welcome opportunities for value enhancement
Nomura believes major global cloud service providers are fully pursuing general artificial intelligence, with the capital expenditure race far from over, so V4 is not expected to cause the kind of shockwave seen in last year's global AI infrastructure market.
However, global large model and application developers are burdened with increasingly heavy capital expenditures.If V4 can significantly reduce training and inference costs while maintaining high performance, it will help these companies turn technology into revenue faster and ease profit pressure.
On the application side, a more powerful and efficient V4 will foster stronger AI agents. The report observes that apps like Ali Tongyi Qianwen already can execute multi-step tasks in a more automated way, with AI agents transitioning from "conversation tools" to "AI assistants" capable of handling complex tasks.
These multitasking agents need to interact with the underlying large model more frequently, consuming more tokens and thus driving up compute demand. Therefore, improvements in model performance will not "kill software", but rather create value for leading software companies. Nomura emphasizes attention to software companies that can take the lead in leveraging next-generation model capabilities to build disruptive AI-native applications or agents. Their growth ceiling may be raised again due to leaps in model capability.
Risk Warning and DisclaimerThe market involves risks; investment should be made cautiously. This article does not constitute personal investment advice and has not considered the individual investment objectives, financial situation, or needs of any particular user. Users should consider whether any opinions, viewpoints, or conclusions in this article fit their specific circumstances. Investment based on this is at your own risk.