DeepSeek R1Open Weight

Released

DeepSeek R1 is an AI model released by DeepSeek on Monday, Jan 20 2025, 25 days after DeepSeek V3. It is an open-weight model — the trained weights are available to download and run. It is a 671B parameterThe learned values that shape how a model responds. B means billion and T means trillion. For mixture-of-experts models, the total count is larger than the number active for each token. Parameter count alone does not measure quality. model with a 128k token context windowHow much information the model can work with in one request, measured in tokens — pieces of words or other data. This includes your question, attachments, conversation and reply. Limits can vary by model variant and provider.. Benchmark results (shown below) cover DROP, SWE-Bench Verified, LiveCodeBench, MATH-500, AIME 2024, GPQA Diamond, and 3 more.

Benchmarks

Coding

SWE-Bench VerifiedCoding — Real coding tasks pulled from open-source projects — the AI has to find and fix actual bugs. A human-checked version of the original SWE-Bench. Higher is better.
49.2%
#70 of 87
LiveCodeBenchCompetitive coding — Coding problems published so recently the AI can't have seen them in training — a contamination-free test of raw programming skill. Higher is better.
65.9%
#29 of 60

Reasoning & science

DROPReasoning — three-shot token F1 on questions requiring reasoning over paragraphs
92.2
#1
MATH-500Reasoning — accuracy on the 500-problem MATH subset
97.3%
#5 of 15
AIME 2024Reasoning — single-answer accuracy on the 2024 invitational mathematics exam without tools
79.8%
#9 of 21
GPQA DiamondScience — Graduate-level science questions in biology, physics, and chemistry — hard enough that subject-matter PhDs score around 65%. Higher is better.
71.5%
#82 of 125
MMLUGeneral knowledge — A 57-subject multiple-choice exam — history, law, medicine, maths — that was the standard measure of how much a model knows from 2020 until roughly 2024, when frontier scores crowded into the high 80s and labs moved on to harder tests. The scores here were published years apart under different testing setups, so read them as a historical record rather than a like-for-like ranking. Higher is better.
90.8%
#2 of 59

Knowledge work

SimpleQAKnowledge — accuracy on short, fact-seeking questions in the original SimpleQA set
30.1%
#6 of 19
MMLU-ProKnowledge — accuracy on the ten-option MMLU-Pro benchmark
84%
#7 of 28

Compare DeepSeek R1 with

DeepSeek R1

About

DeepSeek R1 was released in January 2025 as a reasoning model built on DeepSeek V3 Base. DeepSeek published the main model weights under the MIT License and described a training process combining cold-start supervised data with reinforcement learning. Its research also introduced R1-Zero, an experiment that applied reinforcement learning without preliminary supervised fine-tuning.

R1 and R1-Zero were distinct checkpoints: the final R1 recipe addressed readability and language-mixing problems observed in R1-Zero. The release also included six smaller distilled models based on Qwen and Llama foundations. Those variants made the reasoning work available at different model sizes, with the underlying foundation-model licenses relevant to their use.

Frequently asked questions

DeepSeek R1 was released by DeepSeek on Monday, Jan 20 2025.

All DeepSeek releases

30 tracked