Researchers build an open rival to OpenAI's o1 for under $50
AI researchers at Stanford and the University of Washington just created an open-source reasoning model called s1 for under $50 in cloud compute. By distilling thinking steps from Google's Gemini into a Qwen base model using 1,000 curated questions, s1 matches frontier reasoning models like OpenAI's o1 on math and coding tests.
s1 is an open-source reasoning model built by researchers at Stanford and the University of Washington, designed to replicate o1-style long thinking chains on a shoestring budget.
On benchmarks measuring math and coding, s1 performs similarly to heavyweights like OpenAI’s o1 and DeepSeek’s R1. The kicker is the price tag: the team trained it for under $50 in cloud compute credits, and they released the full model, training code, and dataset on GitHub.
How they pulled it off
Instead of relying on large-scale reinforcement learning like DeepSeek did for R1, the researchers opted for pure supervised fine-tuning (SFT) via distillation.
They started with an off-the-shelf, open-weight Qwen model from Alibaba as the base. Then, they distilled reasoning patterns from Google's Gemini 2.0 Flash Thinking Experimental. They didn't scrape the entire web; the dataset consisted of just 1,000 carefully curated questions paired with step-by-step thinking traces.
The training run took less than 30 minutes on 16 Nvidia H100 GPUs. Niklas Muennighoff, a Stanford researcher on the team, told TechCrunch that renting the necessary compute today would cost roughly $20.
They also used a neat prompt-level mechanism to stretch test-time compute. By literally injecting the word "wait" into the chain of thought, the model was forced to pause, double-check its work, and arrive at more accurate answers.
Where to get it and the catch
Everything is hosted on GitHub for anyone looking to inspect the weights or run local tests. Because it runs on a compact Qwen base, local deployment is accessible without massive hardware rigs.
There is, however, a legal question mark. The dataset draws directly from Google AI Studio's Gemini 2.0 Flash Thinking Experimental, and Google's terms explicitly forbid using outputs to train competing models. Furthermore, distillation replicates capabilities; it doesn't invent new reasoning frontiers that surpass the teacher model.
Why it matters
For independent developers and resource-strapped labs, s1 is a practical playbook. It demonstrates that you do not need millions of dollars to acquire frontier reasoning traits if you have 1,000 pristine examples and a solid open base.
For frontier labs spending billions on infrastructure, it highlights how fragile their defensibility might be. If an o1-level experience can be bottled for $20, the economic moat shifts entirely from raw model smarts to product workflows and infrastructure scale.
Sources (1)
- techcrunch.com(opens in a new tab)Originaltechcrunch.com