DeepSeek-V3: A New Era of Open-Source AI Model

DeepSeek has officially released the first version of its new series model, DeepSeek-V3, and made it open source. This marks a significant milestone in the ongoing evolution of AI technology, offering developers and researchers access to a powerful new tool.

Unleashing the Power of DeepSeek-V3

The DeepSeek-V3 model represents a significant leap forward in AI capabilities. You can interact with the latest V3 model by logging into the official website chat.deepseek.com. The API service has been updated simultaneously, with no configuration changes required for the interface. Note that the current version of DeepSeek-V3 does not support multimodal input and output.

Key Features and Capabilities

• MoE Architecture: DeepSeek-V3 is a self-developed Mixture-of-Experts (MoE) model with 671 billion parameters, activating 37 billion, and pre-trained on 14.8 trillion tokens. (DeepSeek-V3 Paper)
• Performance Benchmarking: DeepSeek-V3 surpasses other open-source models like Qwen2.5-72B and Llama-3.1-405B in several evaluations. It performs comparably with top closed-source models such as GPT-4o and Claude-3.5-Sonnet.

Performance Highlights

• Encyclopedic Knowledge: DeepSeek-V3 shows a substantial improvement in knowledge-based tasks (MMLU, MMLU-Pro, GPQA, SimpleQA), approaching the level of Claude-3.5-Sonnet-1022.
• Long Text Handling: In long text assessments like DROP, FRAMES, and LongBench v2, DeepSeek-V3 outperforms other models on average.
• Coding Prowess: DeepSeek-V3 excels in algorithmic coding scenarios (Code forces), significantly outperforming existing non-o1 models. It also nears Claude-3.5-Sonnet-1022 in engineering coding scenarios (SWE-Bench Verified).
• Mathematical Abilities: DeepSeek-V3 significantly surpasses all other open-source and closed-source models in American Mathematics Competitions (AIME 2024, MATH) and China National High School Mathematics League (CNMO 2024).
• Chinese Language Proficiency: DeepSeek-V3 is on par with Qwen2.5-72B in educational evaluations like C-Eval and pronoun disambiguation benchmarks. It is also more advanced in factual knowledge (C-SimpleQA).

Speed and Efficiency Improvements

One of the most remarkable improvements in DeepSeek-V3 is its generation speed. Through algorithmic and engineering innovations, it has increased from 20 TPS (tokens per second) to 60 TPS, a threefold increase compared to the V2.5 model. This enhancement provides users with a faster and smoother experience.

API Service Pricing

With the launch of the more powerful and faster DeepSeek-V3, the pricing for the model API service has been updated to:

• ¥0.5 per million input tokens (cache hit) / ¥2 per million input tokens (cache miss)
• ¥8 per million output tokens This pricing structure is designed to sustain and improve the provision of high-quality model services. A promotional period offers a discount until February 8, 2025, making it even more accessible for new and existing users.

Open Source and Local Deployment

DeepSeek-V3 is trained using FP8 and open-sources the native FP8 weights. Supported by the open-source community, SGLang and LMDeploy have immediately supported the native FP8 inference of the V3 model. TensorRT-LLM and MindIE have implemented BF16 inference. Additionally, a conversion script from FP8 to BF16 is provided to facilitate community adaptation and expansion of application scenarios. You can download the model weights and find more local deployment information at Hugging Face.

DeepSeek's Commitment to Open Source AGI

DeepSeek's unwavering commitment to open source and long-termism aims to democratize AGI. They are excited to share their progress in model pre-training with the community, observing the narrowing gap between open-source and closed-source models. DeepSeek plans to continue building richer functionalities such as deep thinking and multimodality on the DeepSeek-V3 base model, and to continuously share its latest explorations with the community.

For further information and engagement, you can join the DeepSeek community through:

• Email: (Obfuscated for security - see original article for deobfuscated address)
• Discord: DeepSeek AI Discord
• Twitter: DeepSeek AI Twitter
• GitHub: DeepSeek AI GitHub

Read more about other DeepSeek releases:

• DeepSeek APP Release
• DeepSeek V2 Series Conclusion