DeepSeek R1

DeepSeek R1 is DeepSeek's first reasoning model, released January 2025. It uses chain-of-thought reasoning to solve complex problems, competing directly with OpenAI o1.

Key Specifications

• Architecture: Built on DeepSeek V3 base
• Context window: 128K tokens
• Max output tokens: 128,000 (extended reasoning traces)
• License: MIT (open-source)
• Training: RL with reasoning-focused reward modeling

Training Approach

DeepSeek R1 was trained using a novel pure reinforcement learning approach:

1. Cold-start data — Small amount of supervised reasoning examples
2. RL Stage 1 — Pure RL with reasoning rewards (no SFT)
3. Rejection sampling — Filter low-quality reasoning traces
4. RL Stage 2 — Combined reasoning + helpfulness rewards

This approach demonstrates that reasoning abilities can emerge from RL without extensive supervised data.

Capabilities

• Extended reasoning — Shows step-by-step thinking before final answer
• Self-verification — Checks its own reasoning for errors
• Math & science — Strong at mathematical problem-solving
• Code reasoning — Complex algorithm design and debugging
• Open weights — Available under MIT license

Performance

R1 achieves comparable performance to OpenAI o1 on math (AIME, MATH-500) and coding benchmarks, at significantly lower API cost.

Distilled Versions

DeepSeek also released distilled variants (1.5B-70B) using knowledge distillation from R1, enabling smaller models to gain reasoning capabilities.

Comparison

• DeepSeek V3 — Base model (non-reasoning)
• o1 — OpenAI's reasoning model competitor
• Claude 4 — Anthropic's model with reasoning capabilities
• See reasoning-models-comparison for benchmark comparison