Berkeley drops Sky-T1: an open o1 rival trained for $450

Sky-T1-32B-Preview is a fully open-source reasoning model with an unbeatable pitch: it matches OpenAI's o1-preview in hard math and coding, yet cost less than $450 to train.

Developed by the NovaSky team at UC Berkeley's Sky Computing Lab, the project aims to break down the walls built around reasoning AI. While frontier reasoning models remain black boxes, NovaSky has open-sourced everything: the 32B weights, full training code, evaluation harnesses, and the exact 17,000-sample dataset used during training.

How they pulled it off

Instead of training from scratch with massive compute, NovaSky took an efficient distillation and fine-tuning shortcut. They used Qwen2.5-32B-Instruct as the base model and trained it with Llama-Factory and DeepSpeed Zero-3 on just 8 Nvidia H100 GPUs. The entire run took 19 hours, racking up less than $450 on Lambda Cloud.

The real secret lies in ruthless data curation. The team used QwQ-32B-Preview to generate reasoning traces, then ran rejection sampling to filter out hallucinations. Math solutions had to match ground truth answers exactly, and code snippets had to pass automated unit tests.

Formatting turned out to be equally critical. Reasoning chains generated by raw models often tuck answers into messy spots. By using GPT-4o-mini to standardize the thinking traces, they fixed parsing issues that previously dragged benchmark scores down, like jumping from 25% to over 90% accuracy on the APPs dataset simply because test scripts could finally find the code blocks.

The final training mixture consists of 10,000 math problems, 5,000 coding problems, and 1,000 science puzzles.

Performance and hardware reality

In head-to-head benchmarks, Sky-T1 trades blows with o1-preview. On the demanding AIME 2024 math competition, Sky-T1 scores 43.3%, edging out o1-preview at 40.0% and crushing its base model's 16.7%. On LiveCodeBench-Hard, it hits 17.9% compared to o1-preview's 16.3%.

It still struggles where broad factual reasoning is required. On GPQA-Diamond, a benchmark of graduate-level science questions, Sky-T1 scores 56.8% while o1-preview reaches 75.2%.

Downscaling is also not quite there yet. Experiments on 7B and 14B models resulted in repetitive loops during reasoning. To run or fine-tune this yourself, you will need enough VRAM to host a 32-billion parameter model, but all recipes, weights, and datasets are up on GitHub and Hugging Face ready to use.

Why it matters

Sky-T1 completely demystifies the reasoning capabilities that OpenAI turned into a premium subscription feature. It shows that once a strong open base model exists, turning it into a reasoning powerhouse takes clever data selection and pennies in cloud credits, not millions in venture capital.

For developers and enterprise teams, this opens the door to building domain-specific reasoning agents. Instead of waiting for proprietary APIs to lower rates or grant access to raw reasoning tokens, teams can now adapt NovaSky's exact recipes to train their own in-house problem solvers for a few hundred bucks.