AI News (2026/4/24): DeepSeek Launches DeepSeek-V4 Preview Version
Executive Summary:
DeepSeek has officially launched the preview version of its new model series, DeepSeek-V4, and has also open-sourced it. The series includes deepseek-v4-pro and deepseek-v4-flash, both of which support 1M ultra-long context. V4-Pro matches the performance of top-tier closed-source models in agent encoding, world knowledge, and reasoning, while V4-Flash provides similar reasoning capabilities at a lower cost. The models adopt a new attention mechanism and DSA sparse attention, significantly reducing the computational and memory overhead for long contexts.
DeepSeek Launches DeepSeek-V4 Preview Version
News Details
DeepSeek officially released the preview version of its new model series, DeepSeek-V4, on Friday, April 24, and announced its open-source availability. The series includes DeepSeek-v4-pro and DeepSeek-v4-flash, both of which support 1M ultra-long context. This release marks a significant advancement in DeepSeek's capabilities in the large language model domain, offering more options for developers and researchers.
Key Points
1M Ultra-Long Context: The DeepSeek-V4 series models support 1M ultra-long context, significantly enhancing their ability to handle complex tasks such as long document understanding and multi-turn dialogues.
Agent Encoding and Reasoning Performance: The DeepSeek-v4-pro version achieves the performance level of top-tier closed-source models in agent encoding, world knowledge, and reasoning. This makes it highly effective for tasks requiring advanced intelligence and background knowledge, such as intelligent customer service and virtual assistants.
Cost Optimization: The DeepSeek-v4-flash version provides near-top-tier closed-source model reasoning capabilities at a lower cost through optimized algorithms and architecture design. This is a significant advantage for development teams and enterprises with limited resources.
New Attention Mechanism and DSA Sparse Attention: The DeepSeek-V4 series adopts a new attention mechanism and DSA sparse attention technology, drastically reducing the complexity and memory overhead of long context computations. This makes the models more efficient and scalable in practical applications.
AI-ALL In-Depth Analysis
The release of DeepSeek-V4 not only showcases DeepSeek's technical prowess in the large language model domain but also provides AI developers with more choices. The ability to support 1M ultra-long context gives these models a significant edge in handling complex tasks, especially in applications requiring long-term memory and multi-turn dialogues. The DeepSeek-v4-pro version's performance in agent encoding, world knowledge, and reasoning is on par with top-tier closed-source models, offering powerful support for developers. Meanwhile, the DeepSeek-v4-flash version's cost optimization allows teams with limited resources to enjoy high-performance reasoning capabilities. The application of the new attention mechanism and DSA sparse attention technology further enhances the models' efficiency and scalability, injecting new momentum into the development of the AI ecosystem.
