The Chinese AI company DeepSeek has introduced a new version of its large language model called DeepSeek-V4.1-Flash. This model uses a "Mixture-of-Experts" (MoE) architecture, which allows the system to activate only the necessary parts of the model for specific tasks, improving efficiency. With 552 billion parameters, it is the smallest in DeepSeek's new architecture family and the first to include native visual understanding, meaning it can process images and text together without needing additional tools. The company claims that new training methods and reinforcement learning have made this model outperform earlier versions in benchmark tests. DeepSeek, based in Hangzhou, Zhejiang Province, is a Chinese AI research lab that develops open-source large language models. It is owned by High-Flyer, a Chinese investment fund. The V4.1-Flash model is described as "smarter, faster, and more efficient" due to an asymmetric architecture that uses only 8 billion parameters for processing inputs and 16 billion for generating outputs. This design is said to reduce computational and memory demands significantly, cutting the key-value cache memory requirements to a quarter of previous models and storage needs to an eighth. These improvements are expected to lower costs for users relying on the model. The new model is now available via the DeepSeek API under the name "deepseek-flash," with support for both text and visual data. Older versions, such as V4-Flash and V4-Flash-Vision-Exp, have been retired, though existing API requests will be temporarily redirected to the new model for compatibility. DeepSeek has also reduced its API pricing, maintaining a system where off-peak usage is 50% cheaper than peak hours, encouraging developers to schedule tasks during less busy times. The company plans to work closely with the open-source community and is inviting large-scale users with more than 2,000 GPUs to reach out for deployment support. This announcement comes amid growing concerns from U.S. authorities about Chinese companies potentially extracting data from American AI models. A joint statement claims that companies like DeepSeek, Moonshot AI, Alibaba, and others have used vast numbers of queries to American models like ChatGPT, Gemini, and Claude, spending billions of tokens in the process. The U.S. government has called for a coordinated response to address this issue across the global AI industry.