DeepSeek V3.1 API

DeepSeek-V3.1 is a large hybrid reasoning model that uniquely supports both a "thinking" and a "non-thinking" mode, which can be controlled via prompt templates.

Built upon the DeepSeek-V3 architecture with 37 billion active parameters (out of 671B total), it has been post-trained for an extended 128K long-context window. This version introduces significant improvements in tool calling, agentic tasks, and reasoning efficiency, achieving performance comparable to earlier high-end models but with faster response times. These enhancements make it ideal for complex coding, research, and agentic workflows.