DeepSeek: DeepSeek V3.1

DeepSeek-V3.1 is a large hybrid reasoning model featuring 671 billion parameters (with 37 billion active parameters) that uniquely supports both 'thinking' and 'non-thinking' modes via prompt templates. It builds upon the DeepSeek-V3 base, enhancing its capabilities with a two-phase long-context training process to achieve an impressive context window of up to 128K tokens. For efficient inference, it utilizes FP8 microscaling. Users can precisely control the model's reasoning behavior using the `reasoning` `enabled` boolean parameter. This model significantly improves tool use, code generation, and reasoning efficiency, delivering performance comparable to DeepSeek-R1 on challenging benchmarks while offering faster response times. It supports structured tool calling, advanced code agents, and sophisticated search agents, making it an excellent choice for research, complex coding tasks, and agentic workflows. DeepSeek-V3.1 succeeds the DeepSeek V3-0324 model and demonstrates strong performance across a wide array of tasks. Key specifications include an 8K token context window (with 128K available via long-context training), a maximum output of 4K tokens, and competitive pricing at $0.15/$0.75 per 1M input/output tokens. It offers capabilities such as functions, code generation, and streaming, making it a versatile tool for various AI applications.