LLMOpen weights
DeepSeek-V3
DeepSeek
Updated May 25, 2026 · Benchmarks via Artificial Analysis, specs & pricing via OpenRouter · Methodology · Report an error
Superseded by DeepSeek-V3.1
Specifications
- Type
- LLM
- Access
- Open weights
- Released
- December 26, 2024
- License
- DeepSeek License
- Parameters
- 671B total / 37B active (MoE)
- Context window
- 131,072 tokens
- Knowledge cutoff
- July 31, 2024
- Output speed
- 100 tok/s
- Latency (TTFT)
- 0.5s
- Input
- Text
- Output
- Text
- API pricing
- $0.23 in · $0.91 out / 1M tokens
Capabilities
Function callingStructured output
Benchmarks
Reasoning
| DROP | 91.6 |
| GPQA Diamond | 59.1 |
Coding
General
| MMLU | 88.5 |
| IFEval | 86.1 |
| MMLU-Pro | 75.9 |
| SimpleQA | 24.9 |
| Humanity’s Last Exam | 3.6 |
Long Context
| LongBench-v2 | 48.7 |
Our take
A 671B-parameter mixture-of-experts model trained for a fraction of the usual frontier cost, matching top closed models on many benchmarks. Its efficiency claims reshaped the conversation about training economics.