Model

DeepSeek-R1 671B

OPEN

DeepSeek · DeepSeek-R1 family · released Jan 20, 2025

Open reasoning flagship trained with RL on top of V3-Base; matched OpenAI o1 on math/code and triggered a historic US tech-stock selloff on release.

ReasoningCodingVisionFunction callingTool useAgentic

1574.5

Elo · rank #106

Parameters

671B

Active params

37B (MoE)

Context

128K tokens

Architecture

DeepSeekMoE: 256 routed + 1 shared expert, top-8 routing, 61 layers (3 dense + 58 MoE), Multi-head Latent Attention (MLA)

License

MIT

Languages

—

API price (in/out)

$0.55 / $2.19

Modalities

text

Benchmark results

Bar shows position within the tracked field; marker = field best

Aider PolyglotCoding53.3%#23

best: Claude Opus 4.5 · 89.4%

AIMEMath79.8%#79

best: GPT-5.2 · 100.0%

AlpacaEvalHuman preference87.6%#4

best: Tülu 2+DPO 70B · 95.1%

Arena-HardHuman preference92.3%#5

best: Qwen3-235B-A22B (Non-Thinking) · 96.1%

C-EvalKnowledge91.8%#8

best: Qwen3.6-Plus · 93.3%

CodeforcesCoding2029#19

best: DeepSeek-V4-Pro (Think Max) · 3206

DROPReasoning92.2#3

best: Hunyuan-T1 · 93.1

GAIAAgents30.3%#10

best: MiniMax-M2 · 75.7%

GPQA DiamondReasoning71.5%#115

best: GPT-5.6 · 94.6%

IFEvalReasoning83.3%#79

best: Gemma 4 26B A4B · 98.5%

LiveCodeBenchCoding65.9%#81

best: DeepSeek-V4-Pro (Think Max) · 93.5%

MATH-500Math97.3%#17

best: GPT-5 · 99.4%

MMLU-ProKnowledge84.0%#33

best: Claude Fable 5 · 91.5%

MMLU-ReduxKnowledge92.9%#15

best: Qwen3.7-Max · 95.0%

MMLUKnowledge90.8%#5

best: OpenAI o3 · 92.9%

PutnamBenchMath0.1%#5

best: DeepSeek-Prover-V2 671B · 7.1%

SimpleQAKnowledge30.1%#14

best: GPT-4.5 · 62.5%

SuperGPQAReasoning61.8%#15

best: Qwen3.7-Max · 73.6%

SWE-bench VerifiedCoding49.2%#91

best: Claude Fable 5 · 95.0%

Run it locally

VRAM @ Q4

404 GB

VRAM @ FP16

1342 GB

Fits on (Q4)

Multi-node cluster required

671B MoE far exceeds a single 24GB 4090; Q4_K_M alone needs ~404GB, so real deployments use multi-GPU servers or heavy CPU offload (Unsloth's 1.58-bit dynamic quant fits ~131-170GB for partial-offload rigs).

Quantizations

GGUF Q4_K_M · GGUF IQ1/IQ2 (Unsloth Dynamic) · AWQ · MLX

DeepSeek-R1 family

Elo progression across releases

DeepSeek-R1 671BJan 20251574.5 DeepSeek-R1-Distill-Llama-70BJan 20251384.0 DeepSeek-R1-Distill-Qwen-32BJan 20251009.0 DeepSeek-R1-ZeroJan 20251435.8 DeepSeek-R1-0528May 20251757.4

Resources

HFdeepseek-ai/DeepSeek-R1 GHdeepseek-ai/DeepSeek-R1 DOCSapi-docs.deepseek.com APIAvailable via DeepSeek API

API price $0.55/$2.19 · each benchmark row carries its own source badge (see methodology)