In January, the world witnessed Chinese AI startup Deepseek revolutionizing its cost-effective, cutting-edge AI models. The company has announced two models that rival the performance of the frontier models by Openai and Google: the Deepseek-V3 and the Deepseek-V1. Deepseek paves the way for more careful innovation in AI. Now, a new model has sparked the curiosity of the AI community. Researchers from Stanford and Washington University trained an inference model named S1 for just $50 (approximately Rs 4,400) in cloud computing credits.
What is S1?
Based on research papers, the model S1-32B is an open source, advanced language model focusing on inference tasks. What stands out from other AI models is “test time scaling.” This is a technique that allows you to dynamically iterate the responses using additional computational resources during testing. The S1 reportedly competes directly with Openai’s O1 Reasoning model. This also generates answers to the prompts by considering via related questions, so you can also check your own answers. This method differs from traditional approaches that rely solely on training large-scale language models in advance.
For example, if you encourage your model to explain the cost of replacing your iPhone with an Android tablet, it breaks down the question into several steps. tablet.
How were you trained?
The S1 model is trained by curating a high-quality dataset named S1K. These questions were selected based on difficulty, diversity and quality. The dataset also includes complex mathematics, inference and science problems. Another important aspect of model development is the fine tuning (SFT) monitored on this small dataset. According to the research paper, the SFT required 26 minutes of training on 16 Nvidia H100 GPUs. Regardless of small dataset size, S1 achieved high inference accuracy through the use of knowledge built into a pre-trained base model, QWEN2.5-32B-Instruct.
S1 is also based on a ready-made language model trained to infer by studying Google’s Gemini 2.0 Flash Thinking Experimental questions and answers. Google Models demonstrate the thinking behind any answering process, allowing S1 developers to award small amounts of training data to the model. They essentially taught the S1 model to mimic the Gemini thinking process.
In terms of performance, the S1 is rated on three inference benchmarks: AIME24, Math500, and GPQA diamonds. During testing, the model showed a significant improvement in accuracy, surpassing OpenAI’s closed-source model, O1 preview. The S1 model showed performance improvements of up to 27% in mathematical competition issues. While previous models required reinforcement learning and large datasets, S1-32B showed that effective training using only 1,000 samples could build competitive inference models.
What does AI mean?
The S1 model demonstrates the importance of transparency and open source contributions in AI development. With the development process of S1 now available in public spaces, researchers are hoping for more collaboration and innovation in this field. Researchers also show the need to overcome the limitations of test time scaling, explore alternative budgetary approaches, and suggest the need to further strengthen inference capabilities by applying reinforcement learning techniques. .
The story continues under this ad
In short, the S1 is a groundbreaking model that brings together principles of efficient training, innovative test time scaling, and open source.
©IE Online Media Services Pvt Ltd
Enlarge
