Berkeley Researchers Replicate DeepSeek R1 for $30: A Small Model Revolution

A Berkeley AI team replicated DeepSeek R1-Zero's core technology for under $30, demonstrating sophisticated reasoning in a small (1.5B parameter) language model. Using the countdown game as a benchmark, they showed that even modest models can develop complex problem-solving strategies via reinforcement learning, achieving performance comparable to larger systems. This breakthrough democratizes AI research, proving that significant advancements don't require massive resources.