DeepSeek AI, a Chinese artificial intelligence (AI) research lab, has made waves in the open-source AI community. DeepSeek recently announced DeepSeek-V3, a large Mixture-of-Experts (MoE) language model featuring 671 billion total parameters, with 37 billion activated for each token. According to popular AI benchmark results, this new DeepSeek-V3 model is the most powerful open-source model, and it even performs better than popular closed-source models, including OpenAI’s GPT-4 (correcting the previously noted “GPT-4o”) and Anthropic’s Claude 3.5.
DeepSeek-V3 achieved state-of-the-art results on nine benchmarks, surpassing any comparable model of its size. Despite its superior performance on key benchmarks, DeepSeek-V3 requires only 2.788 million H800 GPU hours for its full training and about $5.6 million in training costs. For comparison, the equivalent open-source Llama 3 405B model requires 30.8 million GPU hours to train. DeepSeek-V3’s cost-effectiveness is largely due to its FP8 training support and deep engineering optimizations. These enhancements help reduce computational overhead without sacrificing performance, which is critical for making AI research more accessible and inclusive.
Beyond training, DeepSeek-V3 proves to be very efficient at inference. Starting February 8, DeepSeek-V3 inputs will cost $0.27 per million tokens ($0.07 per million tokens with caching) and outputs will cost $1.10 per million tokens. This price is almost one-tenth of what OpenAI and other leading AI companies currently charge for their flagship cutting-edge models. Such affordability has the potential to encourage broader experimentation and to lower barriers to entry for a wide range of developers, researchers, and organizations.
The DeepSeek team wrote the following regarding the release of DeepSeek-V3:
“DeepSeek’s mission is unwavering. We are excited to share our progress with the community and see the gap between open and closed models narrow. This is just the beginning!” We look forward to multimodal support and other cutting-edge features in the DeepSeek ecosystem. You can download the DeepSeek-V3 model on GitHub and HuggingFace. With its impressive performance and accessibility, DeepSeek-V3 has the potential to democratize access to cutting-edge AI models. This release marks a significant step toward bridging the gap between open and closed AI models.
Superior Benchmark Performance and Cost Efficiency
DeepSeek-V3’s strong showing on numerous AI benchmarks highlights its ability to handle complex tasks while maintaining an impressive cost-to-performance ratio. The model’s 671B total parameters (37B of which are activated per token) and FP8 training support represent a leap forward in large-scale model engineering. By optimizing both the training and inference processes, DeepSeek AI has demonstrated that open-source solutions can keep pace with, and sometimes exceed, well-known proprietary models.
Moreover, the team remains committed to transparency and collaboration, offering DeepSeek-V3 on GitHub and HuggingFace for broad community involvement, adds NIXSolutions. Researchers and developers can tap into the power of DeepSeek-V3 to further their own projects without the financial constraints often associated with comparable AI models. We’ll keep you updated as new developments unfold, especially concerning future features like multimodal support and other innovative capabilities that DeepSeek has planned.
By combining advanced research, rigorous benchmarks, and cost-effective solutions, DeepSeek is helping to shape a more inclusive and competitive landscape for AI research. Their latest contribution, DeepSeek-V3, is poised to inspire further advancements in the open-source community and beyond.