Overview
NovaSky Team developed and released the Sky-T1, a 32-billion parameter reasoning model adapted from Qwen2.5-32B-Instruct. This model is designed for advanced reasoning, coding, and mathematical tasks, achieving performance comparable to state-of-the-art models like o1-preview while being cost-efficient. Sky-T1 was trained on 17K verified responses from Qwen/QwQ-32B-Preview, with additional science data from the Still-2 dataset, ensuring high-quality and diverse learning sources.
The model supports complex reasoning via long chain-of-thought processes and excels in both coding and mathematical challenges. Utilizing Llama-Factory with DeepSpeed Zero-3 Offload, Sky-T1 training was completed in just 19 hours on 8 H100 GPUs, demonstrating efficient resource utilization. These capabilities make Sky-T1 an exceptional tool for applications in programming, academic research, and reasoning-intensive tasks.
Variants
| No | Variant | Cortex CLI command |
|---|---|---|
| 1 | gguf | cortex run sky-t1 |
Use it with Jan (UI)
- Install Jan using Quickstart
- Use in Jan model Hub:
cortexso/sky-t1
Use it with Cortex (CLI)
- Install Cortex using Quickstart
- Run the model with command:
cortex run sky-t1
Credits
- Author: NovaSky Team
- Converter: Homebrew
- Original License: License
- Papers: Sky-T1: Fully Open-Source Reasoning Model