Chinese startup DeepSeek has created a major stir by releasing its advanced language model, DeepSeek V3, which outperformed giant models like Llama 3.1 and GPT-4o in benchmark tests, all at an unbelievable development cost of just $5.58 million.
Article contents:
- Introduction: A major milestone for China’s AI industry
- Benchmark superiority
- Development on a “joke” budget: How did DeepSeek pull it off?
- Expert praise: A feat achieved with minimal resources
- DeepSeek’s philosophy: Cost-effective development
- Toward democratizing artificial intelligence
- Conclusion: A game-changing new player
- Frequently asked questions
Introduction: A major milestone for China’s AI industry
In a move that represents a qualitative leap for the artificial intelligence industry in China, startup DeepSeek has made headlines by releasing its advanced large language model, DeepSeek V3. Featuring 671 billion parameters, this model not only showed robust performance but also outperformed prominent models from tech giants such as Meta’s Llama 3.1 and OpenAI’s GPT-4o in key benchmark evaluations.
Benchmark superiority
DeepSeek V3 demonstrated superior capabilities in benchmarks evaluating text comprehension, coding, and problem-solving. According to DeepSeek’s technical report, the V3 model not only surpassed models from Meta and Alibaba but also delivered results comparable to top-tier models like OpenAI’s GPT-4o and Amazon-backed Anthropic’s Claude 3.5 Sonnet. This places it in the upper tier of currently available AI models.
Development on a “joke” budget: How did DeepSeek pull it off?
The most astonishing part of this achievement is the unbelievably low cost. The Hangzhou-based company revealed in a WeChat post that DeepSeek V3 was developed at a staggering cost of just $5.58 million, using only 2.78 million GPU hours.
To put this into perspective, Meta’s Llama 3.1 model required approximately 30.8 million GPU hours—more than 10 times what DeepSeek used. The Chinese company managed this by relying on Nvidia’s H800 GPUs, a version specifically designed for the Chinese market, which allowed it to bypass US sanctions blocking access to more advanced chips.
Expert praise: A feat achieved with minimal resources
Computer scientist Andrej Karpathy, one of the most prominent figures in AI, praised the achievement on X (formerly Twitter), noting that DeepSeek managed to create a top-tier model with minimal resources. This proves that innovation in AI does not solely depend on having the largest number of chips or the biggest budget, but also relies on efficiency and a smart approach to training.
DeepSeek’s philosophy: Cost-effective development
DeepSeek, which launched in 2022 out of quantitative firm High-Flyer Quant, emphasizes cost-effective artificial intelligence development. The company’s Fire Flyer GPU clusters proved instrumental in driving innovation. This focus on efficiency could grant it a major competitive edge, particularly in an environment where computing costs are the primary barrier to developing advanced models.
Toward democratizing artificial intelligence
DeepSeek aims to democratize artificial intelligence, offering its models for third-party application development alongside its proprietary chat services. By proving that world-class models can be built at a fraction of the cost, DeepSeek opens the door for more startups and research institutions to compete in this space, potentially accelerating the pace of global innovation.
Conclusion: A game-changing new player
The achievement of DeepSeek V3 represents a pivotal moment. It not only highlights China’s rise as a major AI powerhouse but also challenges the prevailing assumption that developing sophisticated large language models requires billion-dollar budgets. By combining superior performance with remarkable cost efficiency, DeepSeek positions itself as a serious competitor to Silicon Valley tech giants, proving that intelligence can sometimes triumph over brute force in the AI race.
Frequently asked questions
Q1: What is DeepSeek V3?
A: It is an advanced large language model (LLM) featuring 671 billion parameters, developed by Chinese startup DeepSeek.
Q2: How does DeepSeek V3 compare to other models?
A: It has outperformed leading models such as Meta’s Llama 3.1 and OpenAI’s GPT-4o across multiple benchmark tests, placing it among the world’s top models.
Q3: What makes DeepSeek’s achievement so special?
A: The ultra-low development cost. The model was trained for under $6 million, a tiny fraction of the budgets major US corporations spend on their models.
