Stay Ahead of the Curve
Latest AI news, expert analysis, bold opinions, and key trends — delivered to your inbox.
UC Berkeley’s NovaSky Team Pushes Open-Source AI Forward with Sky-T1-32B-Preview
2 min read UC Berkeley's NovaSky team unveils Sky-T1-32B-Preview, an open-source reasoning model rivaling early OpenAI o1. Trained in 19 hours for $450, it excels in math and coding tasks, proving high-level AI can be developed affordably and accessibly. January 13, 2025 15:21
UC Berkeley’s NovaSky team just unveiled Sky-T1-32B-Preview, a fully open-source reasoning model that rivals early versions of OpenAI’s o1—at a fraction of the training time and cost.
Key Details:
-
Built on Open Foundations:
- Sky-T1 is a fine-tuned version of Alibaba’s Qwen2.5-32-Instruct.
- It uses training data generated from QwQ-32B-Preview, an open-source reasoning model.
-
Ultra-Efficient Training:
- Training took just 19 hours on 8 H100 GPUs.
- Total cost: around $450, far below typical AI training budgets.
-
Performance Highlights:
- Sky-T1 matches or exceeds earlier versions of OpenAI’s o1 on several benchmarks, especially in mathematics and coding tasks.
-
True Open Source:
- The entire pipeline—training data, code, and model weights—is fully open-source, offering unprecedented accessibility for developers and researchers.
Why It Matters:
UC Berkeley’s breakthrough proves that high-level reasoning models can be developed efficiently and affordably. This could democratize AI development, enabling smaller labs and organizations to compete with tech giants.
The next wave of AI innovation may come from unexpected places—and open-source could lead the charge.
User Comments (0)
No comments added yet.
Comments will not be approved to be posted if they are SPAM, abusive, off-topic, use profanity, contain a personal attack, or promote hate of any kind.
More »
Editor's Choice
Humva
Repobase
Mootion Storyteller 5
Vsub
Create Studio
ChatGPTOpenAI
Pictory
Midjourney
Writesonic
Jasper
AI Agents