BlockBeats report: On August 14, DeepSeek’s product line received a series of updates, officially launching DeepSeek-V4-Pro and its first Agent product, DeepSeek Harness. Meanwhile, DeepSeek-V4-Pro-0813 has been officially released on Hugging Face, with weights now available for download under the MIT license. The official team has also released deployment guides for vLLM, SGLang, and local execution, enabling developers to deploy this version independently.
Regarding product capabilities, the previously leaked evaluation results for DeepSeek-V4-Pro Agent have been confirmed. The official release of DeepSeek-V4 also includes a price adjustment, with the peak output price for the Pro version increasing by 4.5 times compared to before. Some analysts note that, after the price hike, DeepSeek’s peak Flash pricing now exceeds that of GPT-5.6 Luna, eroding its previous cost advantage.
The Agent direction is the highlight of this update. DeepSeek Harness has been released on npm as DeepSeek’s first Agent product, supporting project management, long-term tasks, multi-Agent collaboration, skills, web search, and context management, and allowing users to create custom Agent presets. Tasks such as coding, research, and writing can be configured with different tools, prompts, and workflows.
Architecturally, DeepSeek Harness adopts the core design principle of “Everything is a plugin,” allowing models, tools, prompts, storage, context, and even the interface to be replaced or combined via plugins. Its underlying framework, Cordis, has publicly released its design paper, decomposing model adapters, tool systems, session logs, and agent loops into interchangeable components that support runtime loading, unloading, and composition.
Market enthusiasm is rapidly heating up. Just one and a half hours after the open-source release of DeepSeek Harness, its GitHub Stars surpassed 22,000, surpassing the previous public records set by xAI Grok-1 and DeepSeek-R1 when they reached 20,000 Stars. Industry focus has shifted from single-model capabilities to the engineering foundations required for Agent runtimes, plugin systems, and recursive self-improvement.