DeepSeek: DeepSeek V3.1

DeepSeek V3.1 is a large-scale hybrid reasoning model with 671B parameters (37B active), capable of operating in both “thinking” and “non-thinking” modes through prompt templates. Building on the DeepSeek-V3 base, it introduces a two-phase long-context training process supporting up to 128K tokens, and leverages FP8 microscaling for more efficient inference. Users can directly control reasoning behavior via a simple boolean toggle.

The model enhances tool use, code generation, and reasoning efficiency, delivering performance on par with DeepSeek-R1 on challenging benchmarks while offering faster response times. With support for structured tool calling, code agents, and search agents, DeepSeek-V3.1 is well-suited for research, programming, and agent-driven workflows. As the successor to DeepSeek-V3-0324, it demonstrates strong performance across a wide range of tasks.