DeepSeek V3.1 is a large hybrid reasoning model with 671B total parameters and 37B active parameters, designed to support both thinking and non-thinking modes within a single architecture. This dual-mode capability lets users toggle between deliberate reasoning and direct response generation depending on the task, enabling a single model to handle everything from quick factual lookups to complex multi-step problem solving. The model extends the V3 base through a two-phase long-context training process that pushes effective context windows to 128K tokens, and FP8 microscaling underpins the inference efficiency to make this scale practical to serve at speed.

Building on the V3 base, V3.1 underwent 840B tokens of continued pretraining specifically for long-context extension before post-training refinements. The post-training phase sharpens agentic capabilities—tool calling, code generation, and multi-step reasoning receive targeted boosts, with measurable gains on SWE-Bench and Terminal-Bench benchmarks. The model ships with strict function calling support and Anthropic API compatibility, making it drop-in ready for agentic pipelines and research workflows. Users can control reasoning behavior via prompt templates, and V3.1 achieves performance comparable to DeepSeek-R1 on difficult benchmarks while delivering answers more quickly, positioning it as a strong choice for coding tasks, research, and complex agent applications.