- Published on
DeepSeek V4.1 Flash Model Officially Released: Outperforms V4 Pro, Features Native Multimodal Visual Understanding, a...
- Authors
- Name
- DeepHorus
DeepSeek officially released its latest AI model, DeepSeek V4.1 Flash, on September 10, 2026. This model is the smallest in its new model architecture series, featuring native multimodal visual understanding capabilities, marking a new breakthrough in efficient, high-performance AI architecture for the company.
The release of DeepSeek V4.1 Flash is aimed at meeting the growing demand for AI applications. Its design goal is to achieve a higher capability ceiling, faster inference speed, greater throughput, and scalability to larger parameter models. As a MoE (Mixture of Experts) model with 552 billion parameters, it adopts an innovative Causal-Encoder-Decoder structure. This design results in asymmetric input and output, with only 8 billion input activation parameters and 16 billion output activation parameters, thereby significantly reducing its cost compared to known models of similar size.
In terms of performance, DeepSeek V4.1 Flash excels. Thanks to a new pre-training method and enhanced reinforcement learning post-training, the model has successfully surpassed the intelligence levels of a range of flagship models, including DeepSeek V4 Pro, in multiple benchmark tests. At the same time, the new generation model drastically reduces the size of the KV Cache. Compared to the previous generation, its demand for HBM (High Bandwidth Memory) is reduced to one-quarter, and its demand for SSD (Solid State Drive) is reduced to one-eighth. This improvement is particularly crucial for lowering the operational costs of Agent-class tasks.
The model's release has had a direct impact on the industry and users. DeepSeek V4.1 Flash has been launched simultaneously on the official API. Users can invoke it simply by changing the model name to deepseek-flash. The older models, V4 Flash and V4 Flash Vision Exp, have been taken offline, but for compatibility, the old model names will be temporarily routed to the new model. Since V4.1 Flash has comprehensively surpassed V4 Pro in performance, cost, speed, and other metrics, the official plan is to orderly discontinue the V4 Pro model and route its requests to the new model, billing at the new unit price.
In terms of commercial partnerships and pricing strategy, Tencent's WorkBuddy, CodeBuddy, and OpenCode, as official partners, have fully integrated DeepSeek V4.1 Flash. Thanks to the cost advantages brought by architectural innovation, the official pricing for V4.1 Flash has been reduced accordingly. The peak-valley pricing mechanism continues to be adopted, with off-peak prices being half of the peak period prices, to encourage users to rationally schedule their task times. The new prices took effect at 12:00 on September 10, 2026.
Looking ahead, the release of DeepSeek V4.1 Flash demonstrates the company's ongoing progress in model efficiency and cost control. Through technological innovation, DeepSeek continuously lowers the barrier to AI applications. The simultaneous release of its open-source models and technical reports also provides important references for industry research and application development. With further model optimization and expanded ecosystem cooperation, this model is expected to play a key role in more practical scenarios.
