DeepSeek R1
Available frontierOpen-source reasoning model rivaling o1
Released January 20, 2025
Last tested: 2026-01-15
About
DeepSeek R1 is a reasoning-focused model that demonstrates chain-of-thought capabilities comparable to OpenAI's o1. It uses reinforcement learning to develop extended reasoning abilities, openly sharing its thinking process.
Capabilities
Pricing
- Input
- $3.00/M per 1M tokens
- Output
- $8.00/M per 1M tokens
Static pricing
Details
DeepSeek R1: Open Reasoning at Scale
DeepSeek R1 represents a landmark achievement in open-source AI - a reasoning model that matches OpenAI’s o1 on many benchmarks while being fully open-weight. The model explicitly shows its thinking process, making it valuable for complex problem-solving tasks.
Reasoning Architecture
R1 was trained using large-scale reinforcement learning without supervised fine-tuning, developing emergent reasoning capabilities:
- Chain-of-Thought: Explicit step-by-step reasoning in responses
- Self-Verification: Model checks its own work during inference
- Extended Thinking: Can spend more tokens on harder problems
Benchmark Performance
DeepSeek R1 achieves remarkable results:
- AIME 2024: 79.8% (comparable to o1)
- MATH-500: 97.3% accuracy
- Codeforces: Rating of 2029 (96.3 percentile)
- GPQA Diamond: Strong scientific reasoning
Open Source Advantage
Unlike proprietary reasoning models, R1 is:
- Fully Open Weight: Available for self-hosting
- Distillable: Smaller R1-distilled models available
- Transparent: Reasoning traces are visible and interpretable
Best Use Cases
- Complex mathematical problem solving
- Multi-step coding challenges
- Scientific and technical analysis
- Tasks requiring verifiable reasoning chains
- Research into reasoning model behavior
Other Together AI Models
See all models- DeepSeek V3
- 671B MoE model with groundbreaking cost efficiency
- Context
- 128K
- Speed
- N/A
- TTFT
- N/A
- Qwen 2.5 72B Instruct
- Alibaba's flagship open-source model
- Context
- 128K
- Speed
- N/A
- TTFT
- N/A
- QwQ 32B Preview
- Open-source reasoning model with thinking capabilities
- Context
- 32K
- Speed
- N/A
- TTFT
- N/A
Get the signal, skip the noise.
Weekly digest of new models and provider updates across 41+ compute providers. Curated for AI builders who ship.