Sky-T1 / SkyThought

The NovaSky team's o1-preview-class 32B reasoning model trained for under $450 (19h on 8×H100, 17K distilled QwQ traces on a Qwen2.5-32B base) — the first fully reproducible open reasoning recipe with data, code, and weights (3.4K★). Follow-ups Sky-T1-32B-Flash (anti-overthinking) and the Sky-T1-7B RL series extended the line and seeded the 2025 open reasoning-RL wave.