DeepSeek-R1: Chinese Open Source Model Surpasses GPT-4o and Llama in Benchmarks
Chinese startup DeepSeek launched the DeepSeek-R1 model on January 20, 2025, made available for free under an open source license. The new AI system outperformed key Western competitors – OpenAI's GPT-4o and Meta's Llama 3.1 – in benchmarks for mathematical reasoning and complex problem-solving.
Top performance at zero cost
According to independent tests released by DeepSeek and confirmed by international media, R1 achieved superior results on metrics such as GSM8K (math problems), MATH, and GPQA (scientific reasoning). The company chose to make the model open, meaning any developer can download, modify, and integrate the system into their projects without paying license fees.
The launch comes amid fierce global competition for AI supremacy. While American giants invest billions in infrastructure and charge subscription or API fees, DeepSeek opts for the opposite strategy: democratizing access to high-performance models to accelerate adoption and innovation.
Impact on market and AI geopolitics
Immediate repercussions were felt in financial markets. NVIDIA shares, the main supplier of chips for AI training, fell about 4% in US trading, reflecting fears that efficient, open models could reduce demand for ultra-powerful hardware. Analysts note that DeepSeek-R1 was trained with significantly fewer computational resources than its competitors, challenging the notion that "more processing power is always better."
From a geopolitical perspective, China's advancement reinforces the narrative that the country is not only keeping pace but beginning to lead in critical areas of artificial intelligence. The launch was widely reported by international media on January 20-21, highlighting the pressure it exerts on American and European companies.
Tech community reactions
AI researchers and engineers celebrated DeepSeek's decision to release the model as open source. "This accelerates research and allows startups and universities worldwide to experiment with cutting-edge technologies without prohibitive costs," commented an MIT data scientist in a post on X (formerly Twitter).
On the other hand, executives from companies like OpenAI and Meta have not yet officially commented. Observers note that if R1 truly outperforms rivals on specific benchmarks, pressure will increase for those models to also become open or reduce their prices.
What to expect
DeepSeek plans to update R1 with new datasets and specialized versions for fields such as medicine, engineering, and programming. Meanwhile, the market closely watches the free model's impact on the sector's competitiveness. If the trend continues, the future of AI may be more open and accessible than previously imagined.
DeepSeek-R1 is already available for download on the company's official website and repositories like Hugging Face.