OpenSuperintelligence
The Most Difficult Project In Human History
Path To Open Superintelligence
A strategic roadmap for building AGI through open collaboration, addressing key challenges and defining our path forward
DeepSeek Sparse Attention - DeepSeek-V3.2-Exp
Advanced research on DeepSeek's innovative sparse attention mechanisms for efficient long-context processing and memory optimization
Tiny Recursive Model
How a 7M parameter model beats 100x bigger models at Sudoku, Mazes, and ARC-AGI using recursive reasoning with a 2-layer transformer
Pretrain LLM with NVFP4
NVIDIA's breakthrough 4-bit training methodology achieving 2-3x speedup and 50% memory reduction without sacrificing model quality
47x Faster Image Generation Training
Diffusion Transformers with Representation Autoencoders achieve state-of-the-art FID 1.13 on ImageNet while training 47x faster (80 vs 1400 epochs)
QeRL: Beyond Efficiency
Quantization-enhanced Reinforcement Learning for LLMs achieves 1.5x speedup and enables RL training of 32B models on a single H100 80GB GPU