DeepSeek-V3.1-Terminus

DeepSeek-V3.2-Exp

Experimental model (September 29, 2025); DeepSeek Sparse Attention (DSA) mechanism; 50-75% lower inference costs; long-context optimization

DeepSeek-V3.2

Production release (December 1, 2025); 671B parameters; DeepSeek Sparse Attention; GPT-5-level reasoning; 128K context

DeepSeek-V3.1

Hybrid model (August 2025); 671B parameters; dual-mode (thinking + non-thinking); 128K context; enhanced tool calling

DeepSeek-V3.2-Speciale

Extended reasoning variant (December 1, 2025); extreme thinking mode; 96% AIME score; gold IMO 2025; outperforms GPT-5-High

DeepSeek-Math

Specialized mathematics model; three variants (Base, Instruct, RL); optimized for STEM problem-solving

DeepSeek-V2

Second generation MoE model; 236B total parameters (21B active); 128K context window; 50% lower training cost

DeepSeek-Coder-V2

Advanced coding model; 236B parameters (21B active); 128K context; 338 programming languages; GPT-4-Turbo-level coding

DeepSeek-V2.5

Refinement of V2; improved training data; enhanced transformer architecture; increased computational power

DeepSeek-Coder

First commercial-grade coding model; 1.3B-33B parameters; supports 80+ programming languages

DeepSeek-LLM

General-purpose foundation model; 67B parameters; outperforms LLaMA-2-70B on reasoning and math

DeepSeek-MoE

Mixture of Experts model; efficient inference; reduced memory and computational requirements

DeepSeek-V3

Third generation foundation; 671B parameters (37B active); 128K context; major leap in capability and reasoning