Can $50 Really Build a DeepSeek R1‑Level Reasoning Model? Inside the s1 Low‑Cost Approach

Core Contributions

The paper attributes the performance of the s1 reasoning model to two technical components: the s1K dataset and a test‑time technique called budget forcing .

s1K Dataset

s1K consists of 1,000 carefully selected problems covering math competition questions, PhD‑level scientific queries, and Olympic‑style challenges. Each entry includes a reasoning trace and a final answer generated by Google’s Gemini Flash‑Thinking model. The data were drawn from existing collections (NuminaMATH, OlympicArena, OmniMath) and supplemented with two proprietary sets (s1‑prob and s1‑teasers). The authors validated the dataset on three criteria: difficulty, diversity, and quality.

Budget Forcing

During inference the authors distinguish two computation modes:

Parallel : subsequent computations run independently (e.g., majority‑vote tasks).

Sequential : later steps depend on earlier reasoning (e.g., long chain‑of‑thought).

The budget‑forcing mechanism controls the amount of test‑time computation for sequential tasks. If the model generates more “thinking” tokens than a preset limit, an end‑of‑thinking token and a final answer are forced, terminating the reasoning phase. To encourage more computation, the end‑of‑thinking token is suppressed and a special Wait token is inserted, prompting the model to continue generating reasoning steps.

Fine‑Tuning Procedure

The final model, s1‑32B , is obtained by supervised fine‑tuning Alibaba’s Qwen2.5‑32B‑Instruct on the s1K dataset. The training used PyTorch Fully‑Sharded Data Parallel (FSDP) on 16 NVIDIA H100 GPUs for 26 minutes, consuming a total of 7 GPU‑hours. The authors equate this compute usage to roughly $20–$50 of cloud rental, explicitly excluding labor, data‑collection costs, and any additional experiments.

Evaluation Results

Benchmarking on two reasoning suites shows:

On AIME 2024 and MATH‑500 , s1‑32B surpasses o1‑preview .

Against the full o1 model and DeepSeek R1 , s1‑32B lags substantially on both test sets.

A second set of scores, obtained by fine‑tuning with alternative data subsets, demonstrates that data selection heavily influences performance.

Clarifications and Misconceptions

The authors label the data‑generation pipeline “distillation,” but the process is more accurately described as synthetic‑data supervised fine‑tuning.

The quoted $50 figure refers only to GPU compute cost for the successful fine‑tuning run, not the total research expenditure.

The model was not trained from scratch; it builds on an existing 32‑B instruction‑tuned base.

Additional experiments beyond the single fine‑tuning run were performed, so the $50 number does not capture the full experimental budget.

Open‑Source Artifacts

All code, data, and model checkpoints are released at https://github.com/simplescaling/s1. A quantized version can be run locally with:

ollama run hf.co/brittlewis12/s1-32B-GGUF:Q4_0

Related Findings (LIMO)

The paper LIMO: Less is More for Reasoning reports that using only 817 high‑quality samples (≈1 % of typical SFT data) raises AIME accuracy from 6.5 % to 57.1 % and MATH accuracy from 59.2 % to 94.8 % when combined with test‑time computation scaling. This result challenges the assumption that massive data volumes are required for strong mathematical reasoning.

Key Takeaway

Careful curation of a small, high‑quality reasoning dataset and the application of budget forcing enable a 32‑B model to achieve competitive scores on specific benchmarks at a very low compute cost, but the approach does not eliminate the large resource gap to state‑of‑the‑art models such as full o1 and DeepSeek R1 .