DeepSeek R1 Shocks the AI World: Matching GPT-4 at 5% of the Cost

What DeepSeek Released

On January 20, 2025, DeepSeek released DeepSeek R1, a reasoning model that immediately shocked the global AI industry. Released under the MIT licence, it is completely free to use, modify, and commercialise. R1 is a 671 billion parameter Mixture-of-Experts model — but MoE architecture only activates a fraction of parameters per token, making it more efficient than its size suggests.

DeepSeek claims R1 was trained for approximately $5.6 million in compute costs — a figure that drew immediate scrutiny given that comparable US frontier models cost hundreds of millions.

Benchmark Performance

On AIME 2024, DeepSeek R1 scored 79.8% vs OpenAI o1's 79.2%. On MATH-500, R1 scored 97.3%, matching o1-preview. On Codeforces competitive programming, R1 reached the 96.3rd percentile — above the vast majority of human competitive programmers. These results used pure reinforcement learning via group relative policy optimisation (GRPO) rather than supervised reasoning traces.

The Market Reaction

Nvidia's stock fell 17% on January 27, 2025 — a single-day loss of approximately $593 billion in market capitalisation, the largest for any company in history. The logic: if frontier AI can be trained for $5.6 million instead of hundreds of millions, demand for Nvidia's H100s would be far lower than projected. US chip export controls, designed to slow China, had instead incentivised more efficient training methods.

Open Source Distilled Models

Alongside R1, DeepSeek released distilled models: R1-Distill-Qwen-1.5B, 7B, 14B, 32B, plus R1-Distill-Llama-8B and 70B. The distilled 7B model outperforms GPT-4o on several reasoning benchmarks — a result that would have been considered impossible six months earlier.

What This Means for Indian Businesses