Tülu 3 70B

Tülu 3 70B is Ai2's 70B post-trained model, released November 22, 2024 alongside an 8B sibling and the paper arXiv 2411.15124. It applies Ai2's open post-training stack, SFT plus DPO plus a novel RLVR reinforcement approach, on top of Llama 3.1, and Ai2 claimed parity with GPT-4-class models on math, instruction following, and chat benchmarks.

What distinguishes it is completeness of openness: weights, training data, code, and evaluation infrastructure are all released under Apache 2.0, making the Tülu 3 recipe one of the most replicated post-training playbooks in the open ecosystem.