DeepSeek计划本周启动其Harness工具的内测阶段[1]。该工具将从小范围用户群组中筛选首批参与者，入选用户需提交个人资料并签署《保密承诺函》[1]。根据保密协议，泄露相关信息将影响用户日后获得DeepSeek各类模型、产品内测以及合作机会[1]。
Harness工具的设计理念借鉴了量化交易系统的思路，将与V4正式版同步发布[1]。该工具具备高性能推理循环、可靠性保障、智能上下文管理、多Agent协调等能力[1]。V4正式版原定于7月中旬上线但尚未发布[1]。根据现有定价，V4 Pro缓存命中的价格为高峰时段每百万token 0.05元、平时0.025元，V4 Flash则为高峰时段0.04元、平时0.02元[1]。V4正式版首次原生支持图像推理[1]。此外，DeepSeek已于7月24日停用了deepseek-chat和deepseek-reasoner两个旧API[1]。
DeepSeek's Harness tool is set to enter beta testing this week, with initial participants selected from a limited user pool and required to sign confidentiality agreements [1]. The platform, which draws architectural inspiration from quantitative trading systems, is designed to support high-performance reasoning loops, reliability assurance, intelligent context management, and multi-agent coordination [1]. Users accepted into the program must submit personal information and sign a non-disclosure agreement; breaches will result in exclusion from future DeepSeek model betas, product trials, and partnership opportunities [1].
Harness is scheduled to launch simultaneously with the official release of V4, marking a significant expansion of DeepSeek's capabilities [1]. The V4 official version was originally planned for mid-July but has been delayed [1]. Once released, V4 will introduce native image reasoning support for the first time and implement a peak-and-off-peak pricing structure, with V4 Pro cache hits priced at 0.05 yuan per million tokens during peak hours and 0.025 yuan during off-peak periods, while V4 Flash cache hits will cost 0.04 yuan and 0.02 yuan respectively [1]. Additionally, DeepSeek discontinued the deepseek-chat and deepseek-reasoner legacy APIs on July 24 [1].
