DeepSeek-V3.1
Advanced reasoning model with hybrid thinking capabilities
About model
Revolutionary Hybrid AI:
DeepSeek-V3.1 is a groundbreaking hybrid model that switches between thinking and non-thinking modes, delivering exceptional performance in reasoning, coding, and agent tasks. Built on a massive 671B parameter MoE architecture with 37B activated parameters, it offers unparalleled flexibility for developers seeking both fast responses and deep analytical capabilities.
Model | FrontierMath Tier 4 | GPQA Diamond | HLE | SciCode | GDPval-AA | Terminal-Bench 2.1 | Agent Arena | FrontierCode | DeepSWE |
|---|---|---|---|---|---|---|---|---|---|
DeepSeek-V3.1 | 74.9% | Related open-source models | Competitor closed-source models | ||||||
 Claude Fable 5 | 87.8% | 92.6% | 53% | 60% | 62% | 85% | 53.5% | 70% | |
Claude Opus 5 | 73.2% | 93.2% | 53% | 56% | 68% | 89% | +16.4pp | 53.4% | 74% |
 GPT-5.6 Sol | 82.9% | 94.1% | 47% | 56% | 61% | 88% | 47.5% | 73% | |
 Grok 4.5 | 24.4% | 93.1% | 40% | 54% | 51% | 82% | +4.2pp | 42.4% | 54% |
GPT-5.6 Luna | 61.0% | 91.1% | 37% | 53% | 54% | 81% | 39.8% | 67% |
Model card
Architecture Overview:
• Mixture-of-Experts (MoE) architecture with 671B total parameters and 37B activated parameters
• Built upon DeepSeek-V3.1-Base with two-phase long context extension approach
• Extended training with 630B tokens for 32K phase and 209B tokens for 128K phase
• Compatible with UE8M0 FP8 scale data format for microscaling optimization
Training Methodology:
• Post-trained on expanded dataset with additional long documents
• 10-fold increase in 32K extension phase training
• 3.3x extension in 128K phase training
• Advanced post-training optimization for tool usage and agent tasks
Performance Characteristics:
• Hybrid mode supporting both thinking and non-thinking operations
• Superior performance on MMLU-Redux (91.8% non-thinking, 93.7% thinking)
• Exceptional coding capabilities with LiveCodeBench Pass@1 of 56.4% (non-thinking) and 74.8% (thinking)
• Advanced math reasoning with AIME 2024 Pass@1 of 66.3% (non-thinking) and 93.1% (thinking)
Applications & use cases
Advanced Reasoning & Analysis:
• Complex mathematical problem solving with step-by-step reasoning
• Scientific research and analysis with transparent thought processes
• Academic writing and research with comprehensive literature review capabilities
• Strategic planning and decision-making with multi-factor analysis
Software Development & Engineering:
• Full-stack application development with multi-language support
• Code review and optimization with detailed explanations
• Architecture design and system planning
• Debugging and troubleshooting with systematic approaches
Agent & Automation Tasks:
• Autonomous code agents for software development workflows
• Search agents for information gathering and analysis
• Multi-step task automation with tool integration
• Workflow orchestration and process optimization
Enterprise & Business Applications:
• Data analysis and reporting with comprehensive insights
• Technical documentation and knowledge management
• Customer support and query resolution
• Process automation and workflow optimization
- TypeLLMChatReasoning
- Main use casesChat
- Fine tuningSupported
- SpeedHigh
- IntelligenceVery High
- DeploymentDedicated
- Parameters671B
- Activated parameters37B
- Context length128K
- Input modalitiesText
- Output modalitiesText
- ReleasedAugust 20, 2025
- Last updatedAugust 26, 2025
- Quantization levelFP4
- External link
- CategoryChat
