DeepSeek更新V4.1-Flash-1017 跑分超Kimi接近Opus且已开源

DeepSeek has recently upgraded its Flash model to the DeepSeek 4.1 Flash 1017 version. Early performance assessments show that this new iteration has outperformed several well-known competing AI models, including Kimi K3 and GLM 5.3, and its overall capability is nearly on par with Opus 5.

To verify the model's practical performance, a targeted test was carried out using the latest version of DeepSeek's harness tool in minimalist mode. The test task was straightforward: generate a Mars landing simulation using HTML and Three.js, requiring strict adherence to physical laws and high visual realism—essentially testing whether a "wish-based" programming approach could deliver a complete and functional result. This prompt has long been used to evaluate new AI models, as it challenges their ability to handle long-term software development tasks in a single attempt.

The test results indicate that the entire task took approximately two and a half hours to complete, costing around two dollars and consuming about 39 million tokens. While the token generation speed was impressive, reaching 250 tokens per second, the model's extended thinking process led to relatively slow overall progress. In terms of performance, the model's 3D modeling ability was notably strong, though there were minor flaws: errors appeared in the parachute modeling, and no audio was integrated into the simulation. Overall, the task completion rate was quite high, and the model's performance in this test was assessed to be very close to that of Kimi K3 Max, which finished a similar task in one and a half hours.

It is also worth mentioning that this new DeepSeek model has been made open source, with its weights available on mainstream AI model sharing platforms.