DeepSeek logo

DeepSeek R1 Zero

DeepSeek

DeepSeek-R1-Zero, a model trained using large-scale reinforcement learning (RL) without a prior supervised fine-tuning (SFT) stage, demonstrated remarkable reasoning performance. Through RL, DeepSeek-R1-Zero naturally developed numerous powerful and interesting reasoning behavioral patterns. However, DeepSeek-R1-Zero faces challenges such as infinite repetitions, poor readability, and language mixing. To address these issues and further improve reasoning performance, we introduce DeepSeek-R1, which incorporates cold-start data before RL. DeepSeek-R1 achieves performance comparable to OpenAI-o1 on math, coding, and reasoning tasks.

Key Specifications

Parameters
671.0B
Context
-
Release Date
January 20, 2025
Average Score
76.5%

Timeline

Key dates in the model's history
Announcement
January 20, 2025
Last Update
July 19, 2025
Today
March 25, 2026

Technical Specifications

Parameters
671.0B
Training Tokens
14.8T tokens
Knowledge Cutoff
-
Family
-
Fine-tuned from
deepseek-v3
Capabilities
MultimodalZeroEval

Benchmark Results

Model performance metrics across various tests and benchmarks

Reasoning

Logical reasoning and analysis
GPQA
Pass@1 Diamond AI: I analysis question with using approach on subtasks. 1: question, in order to understand task. 2: solution on key components and intermediate computation. 3: let's solve each part, steps. 4: all computation on errors. 5: results and final answer. This structured approach allows me errors, maintaining reasoning. When solving complex tasks I I will chain reasoning, in order to was how I to answerSelf-reported
73.3%

Other Tests

Specialized benchmarks
AIME 2024
Cons@64 AI: AIME 2023-11 this and year. I should determine, that this such. AIME — this by mathematics for in American Invitational Mathematics Examination. This level in AMC, which most by mathematics. 2023-11 should on specific test, possible, this 11-question from test AIME 2023 AIME — this test. Each question — this task, answer on which is number from 0 to 999. At is 3 on solution 15 questions. Questions usually require application knowledge by numbers and me need to solve 11-question from AIME 2023Self-reported
86.7%
LiveCodeBench
Pass@1 Metric Pass@1 represents itself proportion tasks, which model can solve with first attempts. When evaluation Pass@1 model receives one attempt for solutions each tasks, and solution with first attempts how result. This metric especially important in contexts, where obtain correct answer with first times, for example, in situations application, when users on accuracy model. Pass@1 can be especially strict metric for complex tasks, since she/it not allows several attempts or This makes her/its base modelSelf-reported
50.0%
MATH-500
# Pass@1 Pass@1 — this metric for evaluation efficiency model at solving tasks, where solution can be She/It represents itself probability obtaining correct solutions for one pass, without repeated attempts. If at us is capability generate k various solutions and their correctness, we we can evaluate Pass@1, at model k solutions, and then how often although would one from these k solutions correct. However at evaluation Pass@1 we understand, probability that, that model correct answer with first attempts, and not with k attempts. Therefore for obtaining evaluation Pass@1 is used : Pass@1 = Pass@k × (1/k) where Pass@k — proportion cases, when although would one from k generated solutions correct. This metric allows evaluate ability model generate correct answers with first attempts, even if for her/its measurement is used generation several answersSelf-reported
95.9%

License & Metadata

License
mit
Announcement Date
January 20, 2025
Last Updated
July 19, 2025

Similar Models

All Models

Recommendations are based on similarity of characteristics: developer organization, multimodality, parameter size, and benchmark performance. Choose a model to compare or go to the full catalog to browse all available AI models.