DeepSeek Models
Chinese innovation • MoE architecture
7 Models
Latest: DeepSeek-R1-0528
DeepSeek-R1-0528
Released May 2025
DeepSeek's upgraded reasoning model with 87.5% AIME accuracy and significantly reduced hallucinations
Parameters
671 billion (37B active)
Context
128,000 tokens
Key Features
Enhanced Reasoning System Prompts JSON Output
+2 more
View Details →
DeepSeek-V3-0324
Released March 2025
DeepSeek's updated V3 model with improved post-training using R1 techniques
Parameters
671 billion (37B active)
Context
128,000 tokens
Key Features
Improved Post-training R1-enhanced Advanced Performance
View Details →
DeepSeek-R1
Released January 2025
DeepSeek's reasoning model competing with OpenAI o1, released under MIT license
Parameters
671 billion (37B active)
Context
128,000 tokens
Key Features
Open Source (MIT) Advanced Reasoning Cost Effective
+1 more
View Details →
DeepSeek-V3
Released December 2024
DeepSeek's latest flagship model with enhanced capabilities and efficiency
Parameters
671 billion (37B active)
Context
128,000 tokens
Key Features
Advanced Architecture Improved Efficiency Enhanced Performance
View Details →
DeepSeek Coder V2
Released June 2024
DeepSeek's specialized coding model with advanced programming capabilities
Parameters
236 billion (21B active)
Context
128,000 tokens
Key Features
Code Specialization Multi-language Repository Understanding
View Details →
DeepSeek-V2
Released May 2024
DeepSeek's flagship MoE model with exceptional efficiency
Parameters
236 billion (21B active)
Context
128,000 tokens
Key Features
Mixture of Experts High Efficiency Open Source
View Details →
DeepSeek Math
Released February 2024
DeepSeek's mathematics-specialized model
Parameters
7 billion
Context
4,096 tokens
Key Features
Math Specialization Problem Solving Step-by-step
View Details →