On October 17, DeepSeek unveiled its latest iteration, version V4.1-Flash-1017. This version places a significant emphasis on bolstering programming, agent, and visual functionalities. It boasts the innovative feature of 'wishful programming,' which allows for the direct creation of complete projects from concise prompts, encompassing areas like 3D modeling and web games.

In terms of performance metrics, this new version outstrips its forerunner across 11 benchmark tests. It not only accomplishes tasks more swiftly but also produces shorter outputs. Moreover, it refines project construction capabilities through self-feedback reinforcement learning techniques.

From a pricing perspective, the per-unit cost remains static. However, through meticulous optimization of the cache hit rate (reaching an impressive 97.9%), the actual expenses are significantly curtailed. To put it into perspective, the average cost to complete 10 programming tasks is a mere 40% of that of Claude Opus 5.5 (0.045 USD compared to 0.1168 USD). Nevertheless, it does take, on average, 2 minutes longer to finish these tasks (2 minutes 38 seconds as opposed to 37 seconds).

Furthermore, DeepSeek has generously released its weights under the MIT license. This move empowers teams to deploy the model independently. Additionally, it has introduced novel, self-developed Search and Fetch API interfaces. These interfaces support web scraping and deliver structured result returns. They are priced at 5 RMB per 1,000 successful requests, with billing set to commence on November 1.

At present, this version is accessible via the official web portal, APP, and API. Developers can seamlessly integrate and utilize it within frameworks such as DeepSeek Harness and Tencent CodeBuddy.