Today, DeepSeek officially announced that DeepSeek-V3.1 has been officially released. According to reports, this upgrade includes the following major changes: hybrid reasoning architecture: one model supports both thinking mode and non-thinking mode; higher thinking efficiency: compared to DeepSeek-R1-0528, DeepSeek-V3.1-Think can give answers in a shorter time; stronger Agent capabilities: through Post-Training optimization, the new model's performance in tool usage and agent tasks has been greatly improved.
The official App and web model have been simultaneously upgraded to DeepSeek-V3.1. Users can freely switch between thinking mode and non-thinking mode through the "Deep Thinking" button.
Programming agent: In the code repair evaluation SWE and the complex task (Terminal-Bench) test in the command line terminal environment, DeepSeek-V3.1 has significantly improved compared to the previous DeepSeek series models.
Search agent: DeepSeek-V3.1 has achieved great improvements in multiple search evaluation indicators. In the complex search test (browsecomp) that requires multi-step reasoning and the multi-disciplinary expert-level difficult test (HLE), the performance of DeepSeek-V3.1 has been significantly ahead of R1-0528.
The test results show that after thought chain compression training, V3.1-Think's average performance on various tasks is the same as R1-0528 while the number of output tokens is reduced by 20%-50%. At the same time, the output length of V3.1 in non-thinking mode has also been effectively controlled. Compared with DeepSeek-V3-0324, it can maintain the same model performance with a significantly reduced output length.
In addition, starting from the early morning of September 6, the DeepSeek open platform API interface call price will be adjusted as follows.
