DeepSeek-v3.1 Release
3 days ago
- #DeepSeek
- #AI
- #MachineLearning
- DeepSeek-V3.1 introduces hybrid inference with Think & Non-Think modes.
- DeepSeek-V3.1-Think provides faster answers compared to DeepSeek-R1-0528.
- Enhanced agent skills with improved tool use and multi-step task handling.
- API updates include separate modes for non-thinking (deepseek-chat) and thinking (deepseek-reasoner).
- Supports 128K context and Anthropic API format.
- Beta API now supports strict function calling.
- Improved performance on SWE / Terminal-Bench and multi-step reasoning tasks.
- V3.1 Base includes 840B tokens of continued pretraining for extended context.
- Updated tokenizer and chat template, with open-source weights available.
- New pricing starts on Sep 5th, 2025, with current pricing until then.