DeepSeek 3.1 Released with Improved Memory and Support for Larger Queries
Credit: DeepSeek
Today, the Chinese company DeepSeek announced an update to its previous artificial intelligence model V3. The new version is already available for testing. The main advantage of the updated model is the increased size of the context window.
DeepSeek V3.1 has 685 billion parameters. The context window now supports 128,000 tokens. That is, the user can enter a query as long as a 400-page book. The accuracy of AI has increased significantly. In the Aider benchmark, which tests the capabilities of neural networks in writing code, the accuracy of DeepSeek V3.1 was 71.4%. This is 1% more than that of the popular AI for programmers – Claude Opus 4. DeepSeek V3.1 also has new tokens that improve search capabilities and the “reasoning” model.
DeepSeek V3.1 is still based on the global R1 model. Now DeepSeek fans are waiting for the release of R2, the successor to R1, and Chinese media blame the delay on the perfectionism of the company’s CEO Liang Wenfeng, problems with the supply of Nvidia accelerators, and the low efficiency of Huawei’s AI accelerators.
