AI laboratory DeepSeek has officially transitioned its flagship model, DeepSeek-V4-Pro (Build 0813), out of preview and into general availability. The system is now globally accessible across web interfaces, mobile apps, and the developer API. This release completes the company's newest product line following the late-July rollout of its open-source companion model, V4-Flash.
Architecturally, the model utilizes a Mixture-of-Experts framework housing 1.6 trillion total parameters. To maintain computational efficiency during inference, only around 49 billion parameters are dynamically activated per token. This design aims to deliver massive model capacity while keeping processing latency and computational demands at manageable levels.
DeepSeek-V4-Pro expands document handling by supporting an input context window of 1 million tokens. Furthermore, the maximum generation limit has been increased to 384,000 output tokens in a single run. These specifications are engineered specifically for enterprise tasks such as exhaustive code repository parsing, multi-case legal reviews, and extensive data synthesis.
For complex problem solving, the system introduces a granular reasoning control mechanism. Developers can choose between three distinct thinking budgets labeled low, high, and max, tailoring latency and cost to specific analytical tasks. The API also incorporates native compatibility with OpenAI's Responses API, streamlining integration into established agent frameworks and coding workflows.
To manage compute loads, DeepSeek is introducing a differentiated pricing structure. Beginning on August 16, developers can access an off-peak pricing tier that provides a 50 percent discount on API usage outside peak operating hours. This approach provides an economical path for batch processing, automated back-office pipelines, and continuous background data evaluations.