💬 The Success of DeepSeek: How China's Open Source Model Challenges ChatGPT
The Chinese company DeepSeek has unveiled a new open-source language model—DeepSeek V3. Unlike many competitors, it is freely available for use, including in commercial projects.
🖥 A Skilled Programmer
DeepSeek V3 has outperformed other large language models (LLMs) in nearly all major benchmarks (see table ⤴️). It particularly excels in programming tasks. In the Aider Polyglot test, which evaluates an AI's ability to write code that can be integrated into existing programs, DeepSeek V3 ranks second only to OpenAI's latest o1 version.
💰 Fast, Affordable, High-Quality
These achievements are backed by rigorous preparation. DeepSeek V3 was trained on a dataset comprising 14.8 trillion tokens. This is equivalent to approximately 11.1 trillion words—akin to the AI reading every book ever written by humanity several times over.
Additionally, DeepSeek V3 is currently the largest open-source model in terms of parameters, boasting 671 billion. The previous leader, Meta's Llama 3.1, has only 405 billion. This massive number of parameters allows the model to better understand context and solve more complex tasks—but it comes at a cost. The model requires extremely powerful hardware to function at acceptable speeds.
On the other hand, DeepSeek has made a breakthrough in efficiency: using specialized Nvidia H800 GPUs, they trained their model in just two months. Moreover, the development of DeepSeek V3 cost the company only $5.5 million. For comparison, OpenAI reportedly spent around $78 million to create GPT-4.
👁 Ideologically Aligned
As a Chinese company, DeepSeek adheres to China's internet authority regulations. Primarily, the company must ensure that its model's responses align with the ruling party's values. As a result, DeepSeek V3 provides vague or neutral answers to sensitive questions about China's political system when asked in its "native" language. However, the chatbot is more willing to discuss complex topics in other languages—though it occasionally chooses to remain silent.
➡️ You can chat with DeepSeek V3 for free here.
More on the topic:
🛑 Top AI Trends for 2024: What's Next?
