A new AI model called s1, outlined in a paper released on Feb. 2, has drawn attention in the AI research community for its cost-effective performance.

The model achieves performance close to state-of-the-art levels while operating on lower costs and simpler infrastructure.

It improves large language models (LLMs) during the inference by extending “thinking time” with simple interventions, like replacing ending tags with prompts such as “Wait.”

Trained on a distilled dataset of 1,000 high-quality examples from Qwen2.5, which was developed by Alibaba Cloud. The s1 model was trained using 16 Nvidia H100 GPUs.

With a single training run lasting 26 minutes, the total computational cost was about US$6.

The cost-efficient nature of s1 allows for more frequent experimentation, even with limited resources.

While larger organizations like OpenAI and Anthropic rely on extensive infrastructure, innovations like s1 show progress can be made within constrained budgets.

However, the release of s1 raises concerns about “distealing,” a practice where models use distilled datasets created by other AI systems.

This issue has led to industry debates, with companies like OpenAI voicing concerns regarding its ethical and legal aspects.