On July 31, the official version of DeepSeek-V4-Flash API quietly went live for public beta testing, marking a key step for this domestic large model company in the Agent field. Unlike the previous "Pro strong, Flash weak" hierarchical logic, after this update, the official version of Flash has approached or even surpassed the V4-Pro preview version from three months ago in multiple Agent benchmark tests.
Official data shows that the official version of V4-Flash scored 82.7 on Terminal Bench 2.1, 76.7 on Cybergym, and 70.3 on Toolathlon verified. Notably, this model has 284 billion parameters, with only 13 billion activated parameters, which is nearly an order of magnitude less than V4-Pro (activated 49 billion). DeepSeek clearly stated that this update involved "only re-training," and the model structure remained completely unchanged. This means that through optimization of training methods and data quality, the lightweight model is narrowing the gap with the flagship model in specific tasks.
The other highlight of this upgrade is the first official unveiling of DeepSeek's self-developed Agent framework, DeepSeek Harness. At the same time, V4-Flash natively supports the Responses API format promoted by OpenAI and is specifically adapted for Codex. Developers can low-costly migrate OpenAI ecosystem Agent applications to the DeepSeek platform.
The strategic intent is very clear: to enter the Agent application market with a high-cost-performance lightweight model. Agent scenarios are extremely sensitive to inference speed and cost; Flash's low price advantage will form a significant competitive edge in multi-turn tasks that repeatedly call tools.
This release came less than two months after DeepSeek completed its first round of financing exceeding 50 billion yuan, and the company's valuation has risen to about 71 billion US dollars. With capital support, DeepSeek is accelerating from "no funding, no listing" to a commercialization sprint. As the official version of Pro is about to arrive, the curtain on the Agent war has already been lifted.
