Sky-T1
32B reasoning model trained for under $450
About
Sky-T1-32B-Preview is a reasoning model from UC Berkeley's Sky Computing Lab, fine-tuned from Qwen2.5-32B-Instruct on 17K curated reasoning traces. Its headline claim is cost: the team says the full training run took 19 hours on 8 H100 GPUs for less than $450, and they published the dataset, code, and training recipe so anyone can reproduce it. On benchmarks it lands near o1-preview on math and code, clearly ahead of its base model. As a research preview it lags o1-preview on science-heavy GPQA-Diamond, and like most reasoning models it is slower to answer than a standard instruct model.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Team-reported: MATH500 82.4, AIME 2024 43.3, LiveCodeBench easy 86.3 / medium 56.8 / hard 17.9, GPQA-Diamond 56.8
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (verified tier, metrics-v1). Evidence: Model card read on Hugging Face (NovaSky-AI/Sky-T1-32B-Preview): 32B from Qwen2.5-32B-Instruct, 17K training examples, <$450 training cost, 19h on 8xH100.; Benchmarks from model card table: MATH500 82.4, AIME2024 43.3, LiveCodeBench splits, GPQA-Diamond 56.8 — team-reported.; Release date Jan 10 2025 confirmed via blog post pubDate and press coverage.; Limitation: GPQA-Diamond trails o1-preview (75.2); preview status.; License: open-source code and data per team; weights openly available..