About Open Superintelligence Lab
We conduct open research on neural networks and LLMs. We bring top-tier company and university experience to everyone to contribute to AI research and accelerate the progress of humanity.
✨Featured Research
DeepSeek Sparse Attention
From O(L²) to O(Lk) - The Lightning Indexer Breakthrough
Our research on DeepSeek's sparse attention mechanism that reduces attention complexity from quadratic to near-linear scaling. We conducted experiments comparing standard attention vs sparse attention, showing up to 302% performance improvements on small LLMs.
Pretrain LLM with NVFP4
4-bit Training Revolution - 2-3x Speedup with 50% Memory Reduction
Our research on NVIDIA's NVFP4 breakthrough in 4-bit floating point training methodology. Through our experiments, we demonstrate 2-3x performance improvements with 50% memory reduction while maintaining model quality comparable to FP8 training on billion-parameter models.
How to Contribute
Choose a Project
Browse our research projects above and select one that aligns with your interests and expertise.
Create GitHub Issues
Create issues to discuss ideas, bugs, improvements, or new research directions. We value meaningful contributions that advance our research goals, not trivial ones like "Fixed typo".
Discuss & Collaborate
Engage in discussions, share research findings, and collaborate with the community to advance the projects.
Contribution Guidelines
Research & Ideation
Share your research ideas, experimental results, and theoretical insights. We welcome contributions from all levels of expertise.
Open Discussion
Use GitHub issues and discussions to engage with the community. Join our Discord server for real-time conversations, ask questions, share knowledge, and help others learn.
Code & Implementation
Contribute code, documentation, and implementations. Follow our coding standards and submit pull requests for review.
Documentation & Learning
Help improve documentation, create tutorials, and develop learning materials to make our research more accessible.