Model Overview
DeepSeek-R1 (deepseek-reasoner) is an advanced reasoning-focused large language model designed specifically for complex problem-solving with explicit chain-of-thought capabilities.
Key Features
- Very high intelligence (4/4 dots rating)
- Medium speed (3/5 lightning bolts rating)
- 64,000 context window
- 8,192 max output tokens
- 32,768 maximum chain-of-thought length
- 2023 knowledge cutoff (estimated)
- Text input support
- Text output support
Technical Specifications
- Pricing:
- Standard hours (08:30-00:30 Beijing Time):
- $0.138 per 1M tokens (input with cache hit)
- $0.551 per 1M tokens (input without cache hit)
- $2.204 per 1M tokens (output)
- Discount hours (00:30-08:30 Beijing Time):
- $0.034 per 1M tokens (input with cache hit)
- $0.138 per 1M tokens (input without cache hit)
- $0.551 per 1M tokens (output)
- Supports: Input: text; Output: text with explicit reasoning steps
- Features: Chain-of-thought reasoning, context caching, step-by-step problem solving
Snapshots
- deepseek-reasoner (current version)
Positioning and Use Cases
DeepSeek-R1 is positioned as a specialized model for tasks requiring advanced reasoning capabilities. It excels at mathematical problem-solving, logical reasoning, complex decision-making, and step-by-step analysis. The model's unique chain-of-thought feature makes it particularly valuable for educational applications, scientific research, financial analysis, and any domain where transparency in reasoning process is critical. Unlike standard LLMs, DeepSeek-R1 explicitly shows its reasoning steps before providing final answers, making it more reliable for high-stakes decision-making scenarios.
Rate Limits
- Information not publicly available
Additional Notes
- Chain-of-thought refers to the model's explicit reasoning process before providing a formal answer
- Output token count includes both chain-of-thought tokens and final answer tokens
- Default maximum output length is 4K tokens if not specified by the user
- The model can generate up to 32K tokens of reasoning steps, making it suitable for extremely complex problems
- Context caching feature helps reduce costs for repeated or similar queries
- When both recharged balance and bonus balance exist, bonus balance will be deducted first
- Particularly well-suited for academic research, scientific computing, and enterprise decision-making
Documentation
Official Documentation
