DeepSeek V3
by DeepSeekopen weightsDeepSeek V3 family
Parameters
671B
Context window
64K tokens
Released
2024-12-26
Input price / 1M tok
$0.27
Output price / 1M tok
$1.10
Modality
tools
About DeepSeek V3
DeepSeek's MoE flagship — 671B total, 37B active per token. Open weights, GPT-4-class benchmarks, and pricing 50-100x cheaper than Western frontier models. Broke the assumption that frontier LLMs need $100M training budgets.
Strengths
- Open weights
- Very cheap on hosted APIs
- MoE efficiency
- Strong code + math
Weaknesses
- Text-only
- 64K context (small vs peers)
- US enterprise wariness of Chinese-origin
Best for
Cost-sensitive productionFine-tuningCode generation