The official version of DeepSeek-V4-Flash is online, with 13 billion activation parameters leveraging the Agent battlefield. On July 31, the official version of the API of DeepSeek-V4-Flash was quietly launched for public testing, marking a key step for this large domestic model company on the Agent track.

The official version of DeepSeek-V4-Flash is online, with 13 billion activation parameters leveraging the Agent battlefield. On July 31, the official version of the API of DeepSeek-V4-Flash was quietly launched for public testing, marking a key step for this large domestic model company on the Agent track. Different from the previous hierarchical logic of "Pro is strong and Flash is weak", after this update, the performance of the official version of Flash in multiple Agent benchmark tests has approached or even surpassed the V4-Pro preview version three months ago. Official data shows that the official version of V4-Flash scored 82.7 points on Terminal Bench2.1, Cybergym scored 76.7 points, and Toolathlon verified scored 70.3 points. It is worth noting that the total parameters of this model are 284 billion, and the activation parameters are only 13 billion, which is nearly an order of magnitude different from V4-Pro (activation 49 billion). DeepSeek made it clear that this update "only re-trained the post-training" and the model structure remained completely unchanged. This means that through optimization of training methods and data quality, lightweight models are closing the gap with flagship models on specific tasks. Another highlight of this upgrade is that the self-developed Agent framework DeepSeek Harness is officially named for the first time. At the same time, V4-Flash natively supports the Responses API format promoted by OpenAI and adapts to the Codex in a targeted manner. Developers can migrate the Agent applications of the OpenAI ecosystem to the DeepSeek platform at low cost. The strategic intention is very clear: to enter the Agent application market with a cost-effective and lightweight model. Agent scenarios are extremely sensitive to inference speed and cost. Flash’s low-cost advantage will form a significant competitive advantage in multiple rounds of tasks where the tool is repeatedly called. This release comes less than two months after DeepSeek completed its first round of financing exceeding 50 billion yuan, and the company’s valuation has climbed to approximately US$71 billion. With capital support, DeepSeek is accelerating its transition from "no financing, no listing" to commercialization. With the official version of Pro coming soon, the Agent War has begun. via AI News (author: AI Base)