DeepSeek V3, a new AI model with 671 billion parameters, has been released and is now available on Together AI. The model, which operates with only 37 billion active parameters, has achieved notable performance, ranking #7 in the Chatbot Arena and being the only open model in the top 10. It supports a full 131K context and includes opt-out privacy controls. DeepSeek V3 runs efficiently, achieving speeds of about 17 tokens per second on 2 M2 Ultras with mlx-lm and mlx.core.distributed. The model's popularity is evident from its download numbers within just 10 days of release. Additionally, DeepSeek V3 can be enhanced with TypingMind, offering features like code and chart visualization, and integration with multiple plugins at a cost that is 94.81% cheaper than GPT-4o. Some users have noted that the excitement around DeepSeek V3 on platforms like LocalLlama/X might be disproportionate given its size of 685 billion parameters.