Today marks the official launch and simultaneous open-sourcing of the DeepSeek-V4 preview edition, which comes in two primary versions: Pro and Flash. DeepSeek-V4 is capable of processing ultra-long contexts of up to one million words, placing it at the forefront in China and the open-source community in terms of agent capabilities, world knowledge, and reasoning performance. Specifically, the V4-Pro stands out as a high-performance flagship model with 49 billion parameters, achieving the best open-source performance in the Agentic Coding evaluation. Its mathematical and coding abilities are on a par with the world's leading closed-source models. On the other hand, the V4-Flash emphasizes a lightweight design and cost-effectiveness, featuring 13 billion parameters. Its reasoning performance closely matches that of the Pro version, offering faster and more economical API services.

From a technical standpoint, the V4 series significantly cuts down on computational and memory demands through the implementation of DSA sparse attention and token dimension compression mechanisms. The API service has also been upgraded concurrently, ensuring compatibility with OpenAI and Anthropic interfaces. Pricing remains budget-friendly, with input costs starting at a mere 0.2 yuan per million tokens for the Flash version and 1 yuan for the Pro version. At present, all products within the Ascend supernode series are compatible with the V4 series, and Huawei Cloud's MaaS platform provides one-click, deployment-free calling services.