DeepSeek AI has officially launched its latest model, DeepSeek-V3, marking a significant milestone in the advancement of open-source artificial intelligence. This comprehensive article explores the key features, performance benchmarks, and accessibility aspects of DeepSeek-V3, providing valuable insights for AI enthusiasts, developers, and industry professionals.
On December 26, 2024, DeepSeek unveiled the first version of its DeepSeek-V3 model series, immediately making it available to the public under an open-source license. This release reinforces DeepSeek's commitment to democratizing AI technology and fostering collaboration within the AI community. Users can start interacting with the DeepSeek-V3 model via chat.deepseek.com, with API services updated accordingly.
DeepSeek-V3 is a self-developed Mixture of Experts (MoE) model, featuring 671B parameters with 37B active parameters, pre-trained on 14.8T tokens. According to the research paper available on GitHub, DeepSeek-V3 outperforms other open-source models like Qwen2.5-72B and Llama-3.1-405B in several benchmarks and rivals top-tier closed-source models like GPT-4o and Claude-3.5-Sonnet.
DeepSeek-V3 introduces significant enhancements in generation speed through algorithmic and engineering innovations. The model's text generation speed has increased from 20 TPS (tokens per second) to 60 TPS, representing a threefold improvement compared to the V2.5 model. This enhancement results in a notably smoother and more responsive user experience.
With the launch of DeepSeek-V3, DeepSeek AI has adjusted its API service pricing to reflect the model's enhanced capabilities. The standard pricing is set at:
To encourage adoption, DeepSeek AI offers a promotional pricing period of 45 days, valid until February 8, 2025. During this period, the API service will be available at the previous rate:
This offer applies to both existing and new users registered before the expiration date. For more details on API services provided by DeepSeek, refer to the DeepSeek API documentation.
DeepSeek-V3 is trained using FP8 and provides natively open-sourced FP8 weights. With support from the open-source community, projects like SGLang and LMDeploy have immediately supported native FP8 inference for the V3 model, while TensorRT-LLM and MindIE offer BF16 inference. To further aid community adaptation and broaden potential applications, DeepSeek AI has provided FP8 to BF16 conversion scripts.
Model weights and additional local deployment details can be found on Hugging Face.
DeepSeek AI continues to demonstrate its dedication to the open-source movement, aiming to make advanced AI technology accessible to everyone. By sharing its latest advancements in model pre-training, DeepSeek AI is actively contributing to narrowing the performance gap between open-source and closed-source models.
The release of DeepSeek-V3 marks a new beginning, with plans to enrich the model with deeper reasoning capabilities and multi-modality features. DeepSeek AI remains committed to sharing its ongoing explorations and advancements with the broader AI community, helping to drive further innovation.
For more information and updates, follow DeepSeek AI through the following channels:
The release of DeepSeek-V3 represents a significant step forward in the field of open-source AI. With its impressive performance, increased generation speed, and accessible pricing, DeepSeek-V3 empowers developers and researchers to explore new frontiers in AI. DeepSeek AI's commitment to openness and collaboration paves the way for continued innovation and democratization within the AI landscape.
