Is the new DeepSeek model coming?

DeepSeek is advancing the grayscale testing of its new model version, potentially serving as the final grayscale version before the official launch of V4.

On February 11, some users received prompts to update the DeepSeek App after opening it. After updating (version 1.7.4), users can experience DeepSeek’s latest model. With this upgrade, the model’s context length has been expanded from 128K to 1M, nearly a tenfold increase; the knowledge base has been updated to May 2025, and multiple core capabilities have been substantially improved.

In internal testing, the author found that DeepSeek, when asked, likely is not V4 in its current version, and is most probably the final evolution of the V3 series or the ultimate grayscale version before the official debut of V4.

Nomura Securities released a report on February 10 stating that the DeepSeek V4 model expected to launch in mid-February 2026 will not trigger the global AI computing power demand panic that occurred with the V3 release last year. The firm believes that the core value of V4 lies in driving AI application commercialization through underlying architecture innovations, rather than disrupting the existing AI value chain.

According to evaluations, the new version’s ability to handle complex tasks has now aligned with mainstream closed-source models like Gemini 3 Pro and K2.5. Nomura further notes that V4 is expected to introduce two innovative technologies, mHC and Engram, which will break through hardware bottlenecks at the algorithm and engineering levels related to computing chips and memory. Preliminary internal tests show that V4’s performance in programming tasks has surpassed contemporaneous models like Anthropic Claude and OpenAI’s GPT series.

The key significance of this release is to further reduce training and inference costs, providing a feasible path for global large language model and AI application companies to alleviate capital expenditure pressures.

Architecture innovations optimized for hardware bottlenecks

Nomura’s report points out that the performance of computing chips and HBM memory bottlenecks are persistent hard constraints for domestic large model industries. The upcoming DeepSeek V4’s introduction of mHC (Meta-Connected Hypergraph Constraint) and Engram architectures are systematic optimizations from both training and inference perspectives targeting these shortcomings.

mHC:

• Full name: “Manifold-Constrained Hypergraph Connection.” It aims to address the bottleneck in information flow and training instability in very deep Transformer models.

• Simply put, it makes the “dialogue” between neural network layers richer and more flexible, while using strict mathematical “guardrails” to prevent information from being amplified or corrupted. Experiments show that models using mHC perform better in mathematical reasoning and similar tasks.

Engram:

• A “conditional memory” module. Its design philosophy is to decouple “memory” from “computation.”

• Static knowledge in the model (such as entities and fixed expressions) is stored in a sparse memory table, which can be placed in inexpensive DRAM. When reasoning, it quickly retrieves from this table. This frees up expensive GPU memory (HBM), allowing it to focus on dynamic computation.

mHC improves training stability and convergence efficiency, partially offsetting the generational gap caused by domestic chips in interconnect bandwidth and computational density; while Engram aims to reconstruct memory scheduling mechanisms, overcoming memory capacity and bandwidth limitations under HBM supply constraints with more efficient access strategies. Nomura believes that these two innovations together form an adaptation scheme tailored for the domestic hardware ecosystem, with clear engineering implementation value.

The report further emphasizes that the most immediate commercial impact of V4’s release is a substantial reduction in training and inference costs. Cost optimization will effectively stimulate downstream application demand, leading to a new cycle of AI infrastructure development. In this process, Chinese AI hardware vendors are expected to benefit from increased demand and upfront investments.

Market landscape shifting from “monopoly” to “multiple contenders”

Nomura’s report reviews the market changes one year after the release of DeepSeek V3/R1. By the end of 2024, DeepSeek’s two models accounted for more than half of the token usage of open-source models on OpenRouter.

However, by mid-2025, as more players entered, its market share had significantly declined. The market has shifted from “single dominance” to “multiple contenders.” The competitive environment faced by V4 is far more complex than a year ago. DeepSeek’s “computing power management efficiency” combined with performance improvements has accelerated the development of large language models and applications in China, also reshaping the global competitive landscape and increasing attention on open-source models.

Opportunities for software companies to increase value

Nomura believes that major global cloud service providers are fully pursuing general AI, and capital expenditure races are far from over, so V4 is unlikely to cause the same level of shock to the global AI infrastructure market as last year.

But worldwide large model and application developers are bearing increasingly heavy capital costs. If V4 can maintain high performance while significantly reducing training and inference costs, it will help these companies convert technology into revenue more quickly, easing profitability pressures.

On the application side, more powerful and efficient V4 will foster more advanced AI agents. The report notes that apps like Alibaba’s Tongyi Qianwen are already capable of executing multi-step tasks more automatically, transforming AI agents from “dialogue tools” into “AI assistants” capable of handling complex tasks.

These multi-tasking intelligent agents require more frequent interactions with underlying large models, consuming more tokens and increasing computing power demands. Therefore, improvements in model efficiency will not “kill software,” but rather create value for leading software companies. Nomura emphasizes the importance of focusing on software firms that can leverage the capabilities of next-generation large models to develop disruptive AI-native applications or agents. Their growth potential could be further elevated by leaps in model capabilities.

Risk warnings and disclaimers

Market risks are present; investments should be cautious. This article does not constitute personal investment advice and does not consider individual users’ specific investment goals, financial situations, or needs. Users should evaluate whether any opinions, viewpoints, or conclusions herein are suitable for their particular circumstances. Investment carries risks, and responsibility rests with the individual.