DeepSeek V4.1 Flash 1017: Improvements in Programming

On October 17, 2026, DeepSeek introduced a new version of its Flash model: DeepSeek-V4.1-Flash-1017. This update promises faster task execution, shorter responses, and improved scores on programming tests. The new functionality is immediately available in the chatbot, the app, and the API, providing seamless integration for users.

The update builds on the previous version V4.1 Flash, which was launched in September. DeepSeek emphasizes that the model can now create fully autonomous projects, ranging from a browser-based racing game to complex 3D models. Users do not need to make adjustments to benefit from these improvements; the new version is delivered automatically through the app or website.

One of the most significant improvements can be found in the Terminal-Bench 4.0, where the model makes a leap from 31.2 to 42.3 points. Additionally, there is an increase in the score on ProgramBench from 20.3 to 29.1. In the programming test DeepSWE, the new model also achieves a score of 77.0, which is better than its competitors such as Claude Opus 5 and GPT-5.6 Sol. However, it is important to note that comparisons have been made with older top models, and independent validation of the results is still lacking.

DeepSeek has also made its Search and Fetch API accessible to developers, allowing them to leverage the company's search engine. The costs are attractive at $0.75 per 1,000 calls. DeepSeek's privacy policy confirms that personal data is stored in China, which remains a concern for users in the Netherlands. Developers running the model on their own servers can avoid this data processing. With this latest update, DeepSeek seems to focus on the demand for speed and cost efficiency for programmers. Whether this new model can sustainably compete with top models like Opus 5.5 and GPT-6 will, of course, depend on future independent evaluations.

Read the full article from AI Insider.