Efficient Post-training
High-quality data, supervised fine-tuning, and preference optimization.
Independent open-source research
Small models.
Bigger possibilities.
Making Small Models Surprisingly Capable.
OpenMiniAI explores how better data, efficient post-training, and reinforcement learning can unlock remarkable capabilities in small language models.
What we're exploring
Small by design. Open by default.
Reinforcement Learning
Learning from verifiable feedback with transparent, reproducible recipes.
Agentic Capabilities
Tool use, multi-step interaction, and learning from environments.
First research project
OpenMiniAI · 0.5B
Our first goal is to study how far a 0.5B language model can go with a complete, reproducible post-training pipeline, from SFT to reinforcement learning and agentic training.