AI reasoning models have often been framed as the product of vast infrastructure, specialized teams, and enormous budgets. A new project from researchers at Stanford and the University of Washington complicates that story: they trained a model called s1 for under $50 in cloud compute credits.
The result is not just another small model release. According to the source article, s1 performs similarly to cutting-edge reasoning models, including OpenAI's o1 and DeepSeek's R1, on tests measuring math and coding abilities. The model, its data, and its training code are available on GitHub.
A small model with a big implication
The s1 project began with an off-the-shelf base model, rather than a new model trained from scratch. The researchers used a small AI model from Qwen, an Alibaba-owned Chinese AI lab, which is available to download for free.
They then fine-tuned it through distillation. In plain terms, distillation means training one model to imitate the answers and reasoning behavior of another model. In this case, the researchers said s1 was distilled from Google's Gemini 2.0 Flash Thinking Experimental.
That matters because reasoning models such as OpenAI's o1 are designed to spend more effort on a problem before answering. The source describes this as test-time scaling, or allowing an AI model to think more before it responds. The s1 researchers were looking for the simplest path to strong reasoning performance and that kind of extended thinking behavior.
The source article says this is related to breakthroughs associated with OpenAI's o1, which DeepSeek and other AI labs have tried to replicate through different techniques. s1 shows that at least some of that behavior can be reproduced with a comparatively lean process.
How the researchers trained s1
The researchers created a dataset of just 1,000 carefully curated questions. Each question was paired with an answer and the associated thinking process from Google's Gemini 2.0 Flash Thinking Experimental.
They used supervised fine-tuning, or SFT, which explicitly instructs an AI model to mimic behaviors found in a dataset. The source contrasts SFT with the large-scale reinforcement learning method DeepSeek used to train R1, its competitor to OpenAI's o1 model.
The training run was short. After less than 30 minutes using 16 Nvidia H100 GPUs, s1 achieved strong performance on certain AI benchmarks, according to the researchers. Niklas Muennighoff, a Stanford researcher who worked on the project, told TechCrunch he could rent the necessary compute today for about $20.
The project also used a simple prompting technique to stretch the model's reasoning. The researchers added the word wait during s1's reasoning, which helped the model double-check its work and arrive at slightly more accurate answers, according to the paper.
Why distillation is becoming a strategic issue
The most important question raised by s1 is not whether it is the strongest model available. The sharper question is economic: if a few researchers can closely replicate parts of a multi-million-dollar model's capability for under $50 in cloud compute credits, what protects the value of the original model?
The source article frames this as a question about the commoditization of AI models. Distillation can make advanced behavior easier to copy, especially when a powerful model is available through an interface that can produce answers and reasoning traces.
That is why large AI labs have reason to be concerned. OpenAI has accused DeepSeek of improperly harvesting data from its API for the purposes of model distillation. The source does not say that accusation relates to s1, but it places s1 in the same broader debate about how model outputs are used to train competitors.
Google also has relevant restrictions. The source says Google offers free access to Gemini 2.0 Flash Thinking Experimental through Google AI Studio, with daily rate limits. It also says Google's terms forbid reverse-engineering its models to develop services that compete with Google's own AI offerings. TechCrunch said it reached out to Google for comment.
What s1 does and does not prove
s1 is a strong signal that reasoning behavior can be transferred more cheaply than many people might expect. Its training cost, dataset size, and release on GitHub make it a useful example for researchers and developers watching the AI reasoning model race.
But the source article also draws an important boundary. Distillation is good at cheaply re-creating an existing AI model's capabilities. It does not, by itself, create new AI models that are vastly better than what is already available.
That distinction helps explain why major AI companies may still invest heavily in infrastructure. In 2025, Meta, Google, and Microsoft plan to invest hundreds of billions of dollars in AI infrastructure, in part to train next-generation AI models. The source notes that such investment may still be necessary to push the envelope of AI innovation.
So the lesson is not that big AI infrastructure no longer matters. The lesson is narrower and more disruptive: once a capability exists, distillation may make it cheaper and faster for others to reproduce pieces of it.
The open-model pressure keeps rising
Because s1 is available on GitHub with its data and code, it adds to the pressure around open AI development. Researchers, startups, and independent developers can inspect the work rather than simply read claims about it.
That transparency is part of what makes the project notable. It gives the AI community a concrete example of a reasoning model trained from a free base model, a small curated dataset, and a short fine-tuning process.
For the biggest AI labs, the project sharpens an uncomfortable reality. Competitive advantage may depend not only on training better models, but also on controlling how model outputs can be used, how access is governed, and how quickly the broader research community can reproduce useful capabilities.
s1 does not end the race for more powerful AI. It does show that the race has more than one track: one for building frontier systems, and another for finding the cheapest practical way to imitate what those systems can already do.