deepseek/deepseek-v3.1

DeepSeek-V3.1 is the flagship language model launched by DeepSeek, primarily designed to mark the first step toward the era of intelligent agents. It features dual-mode reasoning and powerful agent capabilities.

• Dual-mode reasoning: A single model offers both thinking and non-thinking modes, catering to diverse needs.
• Efficient thinking: In thinking mode, it provides faster answers compared to DeepSeek - R1 - 0528.
• API upgrade: Supports the Anthropic API format, with a context window of up to 128K tokens and strict support for Function Calling.
• Improved intelligence: Delivers better performance in programming and search benchmarks, with strong multi-step reasoning capabilities.
• Open-source model: The base model and post-training model weights are available on Huggingface.

───────────────────────────────────────────────────────────────────

Core Capabilities

🧠 Hybrid reasoning capability: Thinking and non-thinking—two modes in one model

⚡️ Faster thinking speed: In DeepSeek-V3.1’s Think mode, the time required is shorter than in DeepSeek-R1-0528 mode, enabling quicker answer generation

🛠️ Enhanced agent capabilities: After post-training, its tool-use abilities and multi-step agent task-processing capabilities have been significantly improved