The new AI model features advanced architecture and prepares for a public offering
Category: Business
On September 10, 2026, Chinese artificial intelligence startup DeepSeek unveiled its latest model, DeepSeek-V4.1-Flash, marking a notable advancement in AI technology. This model, touted as the smallest in its new architecture family, is engineered to deliver superior capabilities, faster inference speeds, and higher throughput, all of which are increasingly important in the rapidly advancing field of artificial intelligence.
DeepSeek V4.1-Flash is a lightweight model that incorporates native multi-modal visual capabilities, allowing it to process various types of data simultaneously. The model employs a cutting-edge 552 billion parameter Mixture of Experts (MoE) architecture, which is complemented by a novel Causal-Encoder-Decoder structure. This innovative design significantly enhances the model’s performance and efficiency.
One of the standout aspects of V4.1-Flash is its asymmetric input and output design. The input activation is set at 8 billion parameters, whereas the output activation is at 16 billion parameters. This unique configuration boosts performance and reduces the operational costs compared to other models of similar size. DeepSeek has also introduced a new pre-training method alongside larger-scale reinforcement learning post-training, which has allowed V4.1-Flash to outperform previous models, including the DeepSeek V4Pro, across various benchmark tests.
Efficiency optimization is a major focus of the V4.1-Flash model. Notably, the size of the KV Cache, which is integral to the model's functioning, has been dramatically reduced. The demand for high-bandwidth memory (HBM) has decreased to one-fourth of what was required by earlier generations, and the need for solid-state drives (SSD) has been cut down to one-eighth. This reduction is particularly beneficial in scenarios where the cost of storing contextual information and cache hits can accumulate significantly.
Statistics reveal that the KV Cache size has been reduced to just 1/437 of the original model's size, a remarkable feat that translates into substantial cost savings for users. With these advancements, DeepSeek aims to provide a more cost-effective solution for AI applications, particularly in high-demand environments.
The launch of DeepSeek V4.1-Flash comes at a strategic time for the company, as it prepares for an initial public offering (IPO) on Shanghai's tech-focused STAR Market. The new model showcases DeepSeek's commitment to innovation and positions the company favorably as it seeks to attract investors. By introducing a model that clearly outperforms its predecessors, DeepSeek is demonstrating its potential for growth and leadership in the competitive AI sector.
DeepSeek has also made systematic adjustments to its product line and billing strategy with the introduction of V4.1-Flash. The previous versions of V4Flash and V4Flash Vision Exp have been officially discontinued, with all requests now routed to the new model. Users can access V4.1-Flash through the DeepSeek API by simply changing the model name to "deepseek-flash." This streamlined integration is expected to facilitate a smoother transition for existing users.
DeepSeek has already secured partnerships with several companies, including WorkBuddy, CodeBuddy, and OpenCode, all of which are under the umbrella of Tencent. These collaborations will enable the partners to fully integrate the new model into their platforms, enhancing their capabilities and offerings in the AI space.
As DeepSeek moves forward, it plans to gradually phase out the V4Pro model. Starting from 12:00 PM on September 14, 2026, requests for the V4Pro will be rerouted to V4.1-Flash, and users will be charged at the new model's unit price. This strategic decision reflects DeepSeek's commitment to consolidating its product offerings around its most advanced technology.
With the launch of V4.1-Flash, DeepSeek is enhancing its technological capabilities and paving the way for future innovations. The company is likely to continue refining its models and exploring new avenues for growth within the AI sector. As the demand for advanced AI solutions continues to rise, DeepSeek's proactive approach may position it as a key player in shaping the future of artificial intelligence.