Tülu 3 405B

Tülu 3 405B applies Ai2's fully open post-training recipe to the Llama 3.1 405B base model, released January 2025. It combines supervised fine-tuning, direct preference optimization, and a novel reinforcement learning method called RLVR (Reinforcement Learning from Verifiable Rewards), which proved especially effective at the 405B scale.

Ai2 reported performance competitive with or superior to DeepSeek V3 and GPT-4o, and ahead of prior open-weight 405B models including Llama 3.1 405B Instruct. As with the smaller Tülu 3 releases, the weights, datasets, training code, and evaluation suite are all open, making it the reference implementation for large-scale open post-training.