DeepSeek于8月13日正式开源DeepSeek-V4-Pro-0813模型，同时推出开源代码智能体框架Harness v0.1开发者预览版，并同步开放配套插件生态系统。[1]
Harness框架采用MIT协议开源，设计理念为"一切皆插件"，对标OpenAI Codex和Anthropic Claude Cowork，旨在为编程和办公场景打造AI生产力工具。[1]新发布的DeepSeek-V4-Pro-0813模型总参数量达1.6万亿，每个token激活490亿参数，支持100万token上下文窗口，最大输出长度可达38.4万token。[1]模型权重采用MIT许可证，支持本地下载自建。[1]
在多项基准测试中，该模型表现出色。在Terminal Bench 2.1测试中得分达87.9分，DeepSWE指标从12.8大幅提升至62.7，DSBench-Hard也提升至67.2。[1]
DeepSeek announced on August 13 the open-source release of its DeepSeek-V4-Pro-0813 model alongside the developer preview of Harness v0.1, an open-source intelligent agent framework designed to rival Anthropic's Claude Cowork and OpenAI's Codex offerings [1]. The framework adopts a "everything-as-plugin" architecture and is licensed under the MIT agreement, enabling developers to build AI productivity tools for programming and office scenarios [1].
The newly released model boasts 1.6 trillion total parameters with 49 billion parameters activated per token, supporting a context window of up to 1 million tokens and maximum output length of 384,000 tokens [1]. In benchmark testing, DeepSeek-V4-Pro-0813 achieved a score of 87.9 on Terminal Bench 2.1, while DeepSWE performance improved from 12.8 to 62.7 and DSBench-Hard reached 67.2 [1]. The model weights are available under an MIT license, allowing users to download and deploy the model locally [1].
