We build open-weight language models that push the frontier of what public research can achieve. Our models are trained on carefully curated datasets and released under permissive licenses, ensuring that the broader research community can build, fine-tune, and deploy without restriction. Every checkpoint is public. Every training run is documented.
Open development is not a marketing strategy. It is the methodology itself. We publish intermediate checkpoints, training logs, and ablation results alongside final models. Researchers can reproduce our work, identify failure modes, and extend our methods without starting from zero. Transparency scales trust.
From reinforcement learning with human feedback to novel architectures for long-context reasoning, our published work spans the full stack of modern language model development. Each paper includes reproducible training configurations and open datasets. We believe that the most impactful research is the kind anyone can verify.