 |
| Photo by Nahi Khabar 24/7 – via Personal Collection |
DeepSeek, the rising Chinese AI startup known for cost-efficient large language models, has announced its latest upgrade — DeepSeek-V3.1. The new model introduces a hybrid inference structure, enabling both reasoning (“think”) mode and non-reasoning mode, marking what the company calls its first major step toward the “agent era.”
According to DeepSeek, V3.1 improves performance in reasoning tasks, speeds up response times compared to its earlier R1 benchmark model, and adds support for agentic capabilities, tool-calling, and enhanced API functionality.
A notable feature is its support for domestic chip compatibility. DeepSeek-V3.1 is optimized for upcoming Chinese domestic chips, using FP8 precision and efficiencies to reduce computational load while maintaining strong performance.
Pricing changes are also part of the package: developers integrating the V3.1 model via DeepSeek’s API will face updated rates starting September 6, 2025. The company says the upgrades and pricing adjustments reflect enhanced capability and cost-effectiveness.
DeepSeek’s continuous upgrades, especially with V3.1, are seen by analysts as part of China’s push to compete globally in AI while managing export restrictions and chip access challenges. With this, DeepSeek hopes to keep pace with larger AI players by offering affordable, capable models.
Post a Comment