Key Specifications
| Vendor | deepseek |
|---|
| Version | v3 |
|---|
| Release Date | 2024-12-26 |
|---|
| Context Window | 64000 tokens |
|---|
| Input Modalities | text |
|---|
| Output Modalities | text |
|---|
| License | DeepSeek License |
|---|
| Documentation | https://api-docs.deepseek.com/ |
|---|
Benchmark Performance
| Benchmark | Score | Unit | Evaluated At | Notes | Source |
|---|
| MMLU | 88.5 | % | 2024-12-26 | 5-shot | view |
| HUMANEVAL | 82.6 | pass@1 | 2024-12-26 | — | view |
| GSM8K | 89.3 | % | 2024-12-26 | 0-shot CoT | view |
| MATH | 61.6 | % | 2024-12-26 | 0-shot CoT | view |
| BBH | 84.9 | % | 2024-12-26 | 3-shot CoT | view |
Compliance
- Data Residency: CN
- SOC2: ✗
- HIPAA: ✗
- GDPR: ✗
- ISO 27001: ✗
DeepSeek V3
Panoramica del modello
DeepSeek V3 是开源 MoE 架构模型，总参数 671B、活跃参数 37B，64K 上下文窗口，在 MMLU、HumanEval、MATH 等基准上达到闭源旗舰水平，价格仅为同级模型的 1/10。
Specifiche principali
| Fornitore | Versione | Data di rilascio | Finestra di contesto | Modalità di input | Modalità di output | Licenza |
|---|
| Deepseek | v3 | 2024-12-26 | 64K | text | text | DeepSeek License |
Prestazioni benchmark
| Benchmark | Punteggio | Unità | Note |
|---|
| MMLU (Massive Multitask Language Understanding) | 88.5 | % | 5-shot |
| HumanEval | 82.6 | pass@1 | — |
| GSM8K (Grade School Math 8K) | 89.3 | % | 0-shot CoT |
| MATH | 61.6 | % | 0-shot CoT |
| BBH (BIG-Bench Hard) | 84.9 | % | 3-shot CoT |
Prezzi
| Input | Output | Lettura cache | Scrittura cache |
|---|
| — | — | — | — |
per milione di token
Punti di forza
- MMLU score 88.5, strong knowledge reasoning.
- HumanEval 82.6, excellent code generation.
- GSM8K 89.3, robust math reasoning.
- 采用 MoE 混合专家架构。
Punti deboli
Casi d’uso
Riferimenti