DeepSeek-V3
Cost-effective high-throughput general reasoning and bilingual generation.
SWE-bench Verified
49.2%
MMLU-Pro
75.9%
MATH-500
90.2%
Context Window
128,000 tokens
| Specification | Value |
|---|---|
| Developer / Lab | DeepSeek-AI |
| Parameters Total | 671B |
| Parameters Active | 37B |
| Architecture Type | DeepSeekMoE + Multi-Head Latent Attention (MLA) |
| License | MIT Permissive |
| Input Pricing (per 1M) | $0.14 |
| Output Pricing (per 1M) | $0.28 |