DeepSeek has officially released DeepSeek V3.1, its most advanced AI model to date. The new model introduces a transformer-based architecture with 560 billion parameters, supporting text, code, and image understanding.
Key improvements include a 43% boost in multi-step reasoning, a 1 million token context window, and a 38% reduction in hallucinations compared to the previous version. The model also supports over 100 languages with near-native proficiency, including better performance in Asian and low-resource languages.
DeepSeek V3.1 is available immediately for enterprise customers via the API, with a gradual rollout to Chrome extension users over the next two weeks. A smaller 7-billion-parameter open-source version has also been released for the research community.
The company highlights applications in software development, scientific research, education, content creation, and business intelligence. DeepSeek emphasizes that V3.1 is a milestone, with ongoing work toward more capable and aligned AI systems.