DeepSeek, a Chinese AI firm, has launched its latest model, DeepSeek V3, which is praised as one of the most powerful open AI models available. Released under a permissive license, this model allows developers to download and modify it for various applications, including commercial use.
DeepSeek V3 excels in handling text-based tasks such as coding, translating, and composing essays or emails from simple prompts. Internal testing indicates that it surpasses both openly available models and closed AI models that require API access.
In competitive coding events on Codeforces, DeepSeek V3 has shown superior performance compared to models like Meta’s Llama 3.1 405B and OpenAI’s GPT-4o. Additionally, it has proven its capabilities in Aider Polyglot, a test measuring how well a model can write new code that integrates with existing code.
DeepSeek claims that their model was trained on an extensive dataset of 14.8 trillion tokens, with 671 billion parameters, making it significantly larger than Llama 3.1 405B. While parameter count often correlates with performance, larger models require more advanced hardware to operate efficiently. DeepSeek V3 was developed using Nvidia H800 GPUs over two months at a relatively low cost of $5.5 million, which is much less than the development expenses for models like GPT-4.
However, the model does have limitations, particularly concerning its responses to politically sensitive topics, as it must comply with Chinese regulations that require it to reflect “core socialist values.”
DeepSeek is backed by High-Flyer Capital Management, which utilizes AI for trading decisions and has invested heavily in AI infrastructure. The founder, Liang Wenfeng, has expressed confidence that closed-source AI models like OpenAI are only a temporary advantage, indicating that competition in the AI space is intensifying.
