DeepSeek has released the V4-Pro with Harness: Flash reaching a peak price higher than GPT-5.6 Luna, while Harness sets a new record for the fastest-growing GitHub Stars.

August 14th - The DeepSeek product line has seen intensive updates, with the official release of DeepSeek-V4-Pro and the launch of the first Agent product, DeepSeek Harness. At the same time, DeepSeek-V4-Pro-0813 has officially integrated with Hugging Face, with weights now available for download under the MIT license. The official release includes vLLM, SGLang, and on-premises deployment instructions, allowing developers to deploy this version independently.

In terms of product capabilities, the previously leaked DeepSeek-V4-Pro Agent evaluation scores have been confirmed. The price of DeepSeek-V4 Pro has been adjusted, with the Pro high-output pricing increasing by 4.5 times. Analysts suggest that post this price increase, the DeepSeek peak Flash price has surpassed that of GPT-5.6 Luna, as the low-price advantage is being reevaluated.

The focus of this update is on the Agent direction. DeepSeek Harness has been published to npm as DeepSeek's first Agent product, supporting project management, long tasks, multiple Agents, Skills, networked search, and context management. Users can create different Agent presets for tasks such as coding, research, and writing, with configurable tools, prompts, and workflows.

Architecturally, DeepSeek Harness adopts the "Everything is a plugin" core design, where models, tools, prompts, storage, context, and even the interface can be replaced or combined through plugins. Its underlying framework, Cordis, has concurrently released a design paper, breaking down model adapters, tool systems, Session Log, and Agent Loop into replaceable components, supporting runtime loading, unloading, and composition.

Market interest has surged rapidly. Just 1.5 hours after DeepSeek Harness was open-sourced, it had exceeded 22,000 GitHub Stars, a growth rate surpassing the previous public records set by xAI Grok-1 and DeepSeek-R1, both crossing the 20,000 Star mark. Industry focus has shifted from singular model capabilities to Agent runtime, a plug-in system, and the engineering foundation required for recursive self-improvement.