News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
S1

S1

A reasoning model using test-time scaling and budget forcing

Released in early 2025 by a research team led by Niklas Muennighoff, s1 is a language model designed to improve reasoning through test-time compute. It uses a technique called budget forcing on a small dataset to outperform larger proprietary models on specific math benchmarks.

Share on WhatsApp

  • Trained on only 1,000 curated examples
  • Uses budget forcing to extend thinking time
  • Outperforms o1-preview on AIME24 math tests
  • Based on the Qwen2.5-32B-Instruct architecture
  • Fully open-source weights and code

What it is

s1 is a reasoning-focused language model released in January 2025. It is built by fine-tuning the Qwen2.5-32B-Instruct model on a curated set of just 1,000 questions. The project aims to demonstrate that strong reasoning performance can be achieved through simple test-time scaling methods rather than massive pre-training datasets. The model is notable for its ability to exceed the performance of larger proprietary systems on competition-level mathematics tasks.

How it came about

The model was developed by a group of researchers including Niklas Muennighoff, Zitong Yang, and Tatsunobi Hashimoto. Their work was published as a paper titled s1: Simple test-time scaling in early 2025. The team sought to replicate the capabilities of OpenAI's o1 model, which had demonstrated advanced reasoning but kept its methods private. By focusing on a minimal dataset and specific inference techniques, they created s1-32B. A successor model, s1.1, was later recommended for better performance.

Inside the model

The architecture relies on supervised fine-tuning of the Qwen2.5-32B base. The training data, known as s1K, consists of 1,000 question-and-reasoning pairs selected for difficulty, diversity, and quality. A key design element is budget forcing, a method to control test-time computation. This technique intervenes during generation by either stopping the thinking process early or extending it by appending the word Wait multiple times. This encourages the model to double-check its logic and correct errors before finalising an answer.

What it can do

On the AIME24 mathematics benchmark, s1-32B scored 56.7 per cent, surpassing the 40.0 per cent achieved by o1-preview. It also reached 93.0 per cent on MATH500. When budget forcing is applied, performance on AIME24 can be extrapolated from 50 per cent to 57 per cent. However, it trails behind newer models like DeepSeek-R1 on some metrics, such as GPQA-Diamond, where it scored 59.6 per cent compared to 71.5 per cent for DeepSeek-R1.

How to use it

The model weights, code, and training data are fully open-source. They are available via Hugging Face and GitHub under the repository simplescaling/s1. Users can run the model locally or through compatible platforms. The developers provide documentation for inference and recommend using the updated s1.1 version for improved results.

Why it matters

s1 demonstrated that high-quality reasoning does not strictly require vast amounts of training data. By showing that a small, carefully curated dataset combined with test-time interventions could rival larger closed models, it influenced subsequent research into efficient reasoning strategies. The open release allowed the community to study and replicate budget forcing techniques.

Questions people ask about S1

What is budget forcing in s1?

Budget forcing is a technique that controls how long the model thinks. It works by either stopping the generation early or adding the word Wait to prompt further reflection. This helps the model verify its answers and fix reasoning errors.

How much data was used to train s1?

The model was fine-tuned on a dataset called s1K, which contains just 1,000 examples. These examples were chosen based on their difficulty, diversity, and quality to maximise learning efficiency.

Is s1 better than OpenAI o1?

s1-32B beats o1-preview on certain math benchmarks like AIME24 and MATH500. However, it does not match the full o1 model or newer open models like DeepSeek-R1 on all tasks, particularly in general science reasoning.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on S1 yet, so the facts on this page come from the sources listed above: arxiv.org/abs/2501.19393, huggingface.co/simplescaling/s1-32B. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.