Introducing an Affordable Alternative: Researchers Develop Open Rival to OpenAI’s o1 ‘Reasoning’ Model for Under $50
AI researchers from Stanford and the University of Washington have successfully trained an AI "reasoning" model known as s1, utilizing less than $50 in cloud compute credits, as detailed in a recent research paper. This model exhibits performance comparable to leading models like OpenAI’s o1 and DeepSeek’s R1 on assessments that evaluate mathematical and coding skills.
The creation of s1 involved fine-tuning an off-the-shelf base model through a process called distillation, wherein reasoning capabilities are extracted by training on an existing AI model’s responses. Specifically, s1 was distilled from Google’s Gemini 2.0 Flash Thinking Experimental model. This method of distilling models has previously been employed by researchers, including those from Berkeley, who created another reasoning model for approximately $450.
The emergence of s1 has sparked discussions about the accessibility of AI model development, raising concerns about how easily others can replicate expensive AI systems with a fraction of the original cost. In response, major AI organizations have expressed their dissatisfaction, with OpenAI accusing DeepSeek of misusing data from its API for distillation purposes.
The goal of the s1 research team was to identify a straightforward approach to achieving strong reasoning performance along with “test-time scaling,” which allows an AI to take more time before responding. Their findings suggest that reasoning models can be effectively distilled even with a limited dataset through supervised fine-tuning (SFT), which is a more cost-effective method than the large-scale reinforcement learning typically used by rivals.
The training process for s1 involved a meticulously curated dataset of just 1,000 questions and corresponding answers from Google’s Gemini model, and the training itself took less than 30 minutes using 16 Nvidia H100 GPUs. The costs incurred for computing resources were approximately $20, indicating a remarkable efficiency in the model’s development.
S1 also incorporates an interesting technique to enhance accuracy—by prompting the model to "wait" before providing answers, researchers observed improved reasoning performance. As major companies like Meta, Google, and Microsoft plan significant investments in AI infrastructure, the capabilities of models like s1 could prompt further exploration into the affordability and commoditization of AI technologies in the coming years.
For further details on the research, the relevant research paper and the s1 model source code are available online.
Discover the pinnacle of WordPress auto blogging technology with AutomationTools.AI. Harnessing the power of cutting-edge AI algorithms, AutomationTools.AI emerges as the foremost solution for effortlessly curating content from RSS feeds directly to your WordPress platform. Say goodbye to manual content curation and hello to seamless automation, as this innovative tool streamlines the process, saving you time and effort. Stay ahead of the curve in content management and elevate your WordPress website with AutomationTools.AI—the ultimate choice for efficient, dynamic, and hassle-free auto blogging. Learn More
