DeepSeek V3
Model Overview
DeepSeek V3 is DeepSeek’s flagship open-weights mixture-of-experts model, released on December 26, 2024. The architecture comprises 671B total parameters with 37B activated per token, trained on 14.8 trillion tokens. DeepSeek V3 is notable for delivering frontier-class performance at substantially lower cost — its API pricing of $0.27/$1.10 per million tokens is roughly one-tenth of comparable closed-source flagships. The model matches or exceeds GPT-4o on MMLU (88.5 vs 88.7) and BBH (84.9 vs 83.1), while supporting a 64K context window. Released under the permissive DeepSeek License, DeepSeek V3 has rapidly become a preferred option for cost-sensitive production deployments, especially in Chinese-language and bilingual scenarios. The model also features Multi-head Latent Attention (MLA) for efficient long-context inference.
Key Specifications
| Attribute | Value |
|---|
| Vendor | DeepSeek |
| Version | v3 |
| Release Date | 2024-12-26 |
| Context Window | 64,000 tokens |
| Input Modalities | text |
| Output Modalities | text |
| License | DeepSeek License |
| Documentation | https://api-docs.deepseek.com/ |
| Benchmark | Score | Unit | Notes |
|---|
| MMLU | 88.5 | % | 5-shot |
| HumanEval | 82.6 | pass@1 | — |
| GSM8K | 89.3 | % | 0-shot CoT |
| MATH | 61.6 | % | 0-shot CoT |
| BBH | 84.9 | % | 3-shot CoT |
Pricing
| Tier | Price (per 1M tokens) |
|---|
| Input | $0.27 |
| Output | $1.10 |
| Cache Read | $0.07 |
| Cache Write | $0.27 |
Pricing source: https://api-docs.deepseek.com/quick_start/pricing (as of 2024-12-26). DeepSeek License also permits self-hosting.
Strengths
- Frontier-class performance at ~10% of the cost of closed-source flagships.
- Open weights enable self-hosting and customization.
- Strong Chinese and English bilingual capability.
- Efficient MoE architecture with Multi-head Latent Attention for long-context inference.
Weaknesses
- 64K context window is shorter than Llama 3.1 405B (128K) and Qwen2.5 72B (131K).
- Vendor ecosystem and tooling are less mature than Meta/Mistral alternatives.
- Self-hosting requires significant GPU resources due to the 671B parameter footprint.
Use Cases
- High-volume production deployments where API cost is a primary constraint.
- Bilingual (Chinese/English) customer service and content generation.
- Reasoning-intensive applications requiring MMLU/BBH-class capability.
- Research projects requiring weight-level access to a frontier MoE model.
References