February 6, 2025:
New $50 LLM Outperforms OpenAIs Model - Researchers from Stanford and the University of Washington have created s1-32B, a large language model that surpasses OpenAI's o1-preview in certain tasks at a cost below $50. The model optimizes reasoning efficiency using test-time scaling and a budget forcing technique to correct potential errors.
Derived from Alibaba's Qwen2.5-32B-Instruct, s1-32B was trained with advanced AI-generated summaries. It notably outperformed o1-preview in MATH and AIME24 benchmarks, requiring only $20 in hardware for training.
