Researchers at the Allen Institute for AI released Tulu 3, a family of language models built on Meta's Llama 3.1 and refined afterwards using openly published data, code and step-by-step recipes. The team reports the models score better than the instruction-tuned versions of Llama 3.1, Qwen 2.5 and Mistral, and better than GPT-4o-mini and Claude 3.5 Haiku, on its own evaluation suite. The release includes a new training method that rewards answers which can be automatically checked for correctness, plus the evaluation toolkit and a report on approaches that did not work.
What changed
Published recipes and training data for refining language models after their initial training were far less detailed than those held privately by commercial labs.
What it unlocks
Reproducing a full post-training pipeline, including the datasets, evaluation toolkit and training code, and adapting it to other domains.
- Tulu 3 405B results added in latest revision
- v1 submitted 22 November 2024, v5 revised 14 April 2025
What you need to act on it
- substantial compute to train models
- familiarity with model fine-tuning
- arxiv.org2026-08-10