DeepSeek-R1

DeepSeek-R1 is a first-generation reasoning model developed by DeepSeek-AI, designed to excel in complex problem-solving. It builds upon the foundation of the DeepSeek-V3-Base model and incorporates advancements in reinforcement learning (RL). The model comes in several versions, including DeepSeek-R1-Zero and various distilled models.

Key Features of DeepSeek-R1

Advanced Reasoning: The model uses a unique training pipeline combining reinforcement learning and supervised fine-tuning to achieve high performance in reasoning, math, and code-related tasks.

Reinforcement Learning: DeepSeek-R1-Zero was trained using large-scale reinforcement learning without supervised fine-tuning, enabling self-verification, reflection, and long chain-of-thought reasoning.

Cold-Start Data: To address issues like repetition, readability, and language mixing in DeepSeek-R1-Zero, DeepSeek-R1 incorporates cold-start data prior to RL training.

Distillation: The reasoning capabilities have been successfully transferred into smaller models while maintaining high performance.

Open Source: The base models and six dense distilled models based on Llama and Qwen are open-sourced for research.

Performance: DeepSeek-R1 achieves performance comparable to OpenAI's models across various benchmarks, with some distilled models outperforming OpenAI-o1-mini.

Parameters: 671B total with 37B activated parameters

Context Length: 128K

Performance Highlights of DeepSeek-R1

• Outperforms several models in English, code, math, and Chinese benchmarks

• Achieves top scores in MMLU-Redux, DROP, AlpacaEval2.0, ArenaHard, Codeforces, and AIME 2024

• DeepSeek-R1-Distill-Qwen-32B sets new state-of-the-art results for dense models.