Llama 3.1 Nemotron 70BOpen Weight

Llama 3.1 Nemotron 70B, released October 15, 2024, was a post-trained build of Meta's Llama 3.1 70B, aligned with NVIDIA's HelpSteer2 preference data and a REINFORCE-based RLHF recipe. For a few weeks it was the open-weight model everyone was talking about: on the preference-style evaluations NVIDIA published — Arena Hard, AlpacaEval 2 and MT-Bench — it edged out GPT-4o and Claude 3.5 Sonnet, the two strongest proprietary models of the moment.

The headline was less the scores than what they demonstrated: an alignment recipe good enough to push someone else's base model past the frontier labs on human-preference benchmarks, published with the training data alongside the weights. It previewed the division of labour NVIDIA formalised the following spring with the Llama Nemotron reasoning family — Meta's pre-training, NVIDIA's post-training, open weights at the end.