USOrigin: USOpen weights
Tülu 3 (Ai2)
Ai2’s open post-training recipe—instruction-following done in the open.
Best for
Students studying alignment/post-training and open instruct checkpoints.
Tip for students
Read the Tülu data/recipe notes before claiming you ‘fine-tuned like Ai2’—the mix matters as much as the base model.
Also in this category
- Ai2 Embodied AIAi2 research on robots, 3D reasoning, and simulation.
- Allen Institute for AI (Ai2)Seattle’s open-AI institute—models, science agents, and planetary tools.
- AstaBench (Ai2)Open benchmarks for scientific AI agents—measure before you trust.
- Hugging FaceThe town square for models, datasets, and demos.
Comparisons
What it is
Tülu is how Ai2 teaches the community post-training (SFT, preference optimization, and related recipes) with full artifacts—not a mysterious closed fine-tune. Often discussed alongside Olmo base models. Ideal when your course asks ‘how do instruct models get good?’
New to the vocabulary? Start with Learn, compare options in Comparisons, or return to the tools index.
ai2 · tulu · post-training · instruct · open-weights