Model
Model explorer

DeepSeek-V3 671B

OPEN
DeepSeek · DeepSeek-V3 family · released Dec 26, 2024

671B/37B MoE trained on 14.8T tokens for ~$5.6M (2.788M H800 GPU-hours); matched GPT-4o/Claude 3.5 quality at a fraction of the cost.

ReasoningCodingVisionFunction callingTool useAgentic
1269.5
Elo · rank #163
Parameters
671B
Active params
37B (MoE)
Context
128K tokens
Architecture
MoE transformer, 671B total / 37B active — DeepSeekMoE + MLA + multi-token prediction, auxiliary-loss-free load balancing
License
DeepSeek Model License (code: MIT)
Languages
API price (in/out)
$0.2 / $0.8
Modalities
text
Benchmark results
Bar shows position within the tracked field; marker = field best
Aider PolyglotCoding49.6%#26
best: Claude Opus 4.5 · 89.4%
AIMEMath39.2%#127
best: GPT-5.2 · 100.0%
AlpacaEvalHuman preference70.0%#9
best: Tülu 2+DPO 70B · 95.1%
Arena-HardHuman preference85.5%#21
best: Qwen3-235B-A22B (Non-Thinking) · 96.1%
C-EvalKnowledge86.5%#21
best: Qwen3.6-Plus · 93.3%
DROPReasoning91.6#4
best: Hunyuan-T1 · 93.1
GPQA DiamondReasoning59.1%#168
best: GPT-5.6 · 94.6%
HumanEvalCoding82.6%#47
best: Claude Opus 4.5 · 99.4%
IFEvalReasoning86.1%#49
best: Gemma 4 26B A4B · 98.5%
LiveCodeBenchCoding37.6%#136
best: DeepSeek-V4-Pro (Think Max) · 93.5%
MATH-500Math90.2%#46
best: GPT-5 · 99.4%
MMLU-ProKnowledge75.9%#71
best: Claude Fable 5 · 91.5%
MMLU-ReduxKnowledge89.1%#26
best: Qwen3.7-Max · 95.0%
MMLUKnowledge88.5%#22
best: OpenAI o3 · 92.9%
SimpleQAKnowledge24.9%#19
best: GPT-4.5 · 62.5%
SWE-bench VerifiedCoding42.0%#100
best: Claude Fable 5 · 95.0%
Run it locally
VRAM @ Q4
400 GB
VRAM @ FP16
1342 GB
Fits on (Q4)
Multi-node cluster required
671B total params vastly exceed a single RTX 4090's 24GB even at 4-bit; requires a multi-GPU cluster or heavy CPU/RAM offload
Quantizations
GGUF · MLX
API price $0.2/$0.8 · each benchmark row carries its own source badge (see methodology)