Download statistics from platforms like HuggingFace, GitHub, etc.

DeepSeek Releases Open-Source DeepSeek-V3 Model; Surpasses Llama 3.1 405b, Offers Cheaper API Access

Dec 26, 2024, 11:51 AM

DeepSeek has unveiled DeepSeek-V3, a new open-source language model with 671 billion parameters utilizing a Mixture of Experts (MoE) architecture with 256 experts and 8 activated per token. The model surpasses Llama 3.1 405b, Claude Sonnet 3.5, and GPT-4o on various benchmarks, including ranking first on BigCodeBench-Hard with an average score of 34.5% and achieving a 60.4 LiveBench score. DeepSeek-V3 was trained on 14.8 trillion tokens at a cost of approximately $5.6 million, significantly less than comparable models. The model delivers enhanced capabilities at 60 tokens per second, three times faster than its predecessor. Additionally, DeepSeek-V3 API is 250% cheaper than Sonnet 3.5, priced at $0.27 per million input tokens and $1.10 per million output tokens. DeepSeek-V3 is fully open-source and available on platforms such as HuggingFace, with API compatibility intact. The release marks a significant advancement in open-source AI, demonstrating that high-performing models can be developed with limited compute resources.