Open-Source Replication of OpenAI’s o1 Model Achieves Superior Performance with Minimal Cost

A paper released by Fei‑Fei Li’s research group demonstrates that the state‑of‑the‑art reasoning capabilities of OpenAI’s o1 model can be reproduced with dramatically lower resources by fine‑tuning the open‑source Qwen2.5‑32B‑Instruct model.

The team performed supervised fine‑tuning (SFT) on a tiny, manually curated dataset of 1,000 high‑quality examples, completing training in just 26 minutes on 16 H100 GPUs, with an estimated compute‑rental cost of about US$20.

Experimental results reveal that the resulting s1‑32B model outperforms o1‑preview by up to 27 % on mathematics competition benchmarks such as MATH and AIME‑24, despite the modest training budget.

The full training pipeline and model weights have been open‑sourced at https://github.com/simplescaling/s1 .

The authors also introduce a “budget‑forcing” technique that deliberately extends inference time by inserting repeated “Wait” tokens or forcibly terminating generation, prompting the model to self‑check and often correct erroneous reasoning steps.

An illustrative example shows the model initially producing an incorrect answer “2”; after the forced interruption and additional “Wait”, the model revises its answer, mimicking a human reviewer’s correction.

Compared with large‑scale reinforcement‑learning approaches (e.g., DeepSeek R1, which requires millions of samples), SFT‑based distillation is far cheaper and more data‑efficient, enabling high‑performance models to be trained on limited hardware.

The authors conclude that while distillation cannot create fundamentally new capabilities beyond current models, it offers a practical path for academic and small‑scale teams to achieve breakthroughs without multi‑million‑dollar budgets.