Meta AIRA₃ wins gold medal at NVIDIA Kaggle competition, ranks 8th out of ~3800 teams
On Sep 5 Meta disclosed details of AIRA₃ autonomous research agent, which won gold at an NVIDIA Kaggle competition to improve reasoning in a 30B Nemotron model, ranking 8th overall. The system uses multiple long-running agents that coordinate asynchronously via a shared forum and filesystem.
On September 5 Meta published detailed architecture notes for its AIRA₃ autonomous research agent. The system won a gold medal in an NVIDIA Kaggle competition focused on improving reasoning in a 30B Nemotron model, ranking 8th overall out of approximately 3,800 teams. AIRA₃ uses multiple long-running agents that coordinate asynchronously via a shared forum and filesystem inside isolated environments.
AIRA₃ outperformed prior agents by roughly 9 percentage points on multiple ML benchmarks. The methodology has reference value for research teams working in machine learning and the sciences, but has no immediate impact on public AI tools. Meta disclosed on the same day that Muse Spark 1.3 shares some of the training infrastructure with AIRA₃. Muse Spark 1.3 is the flagship Muse Code model released on September 2 and was scored at 61 on Artificial Analysis, up 8 points from version 1.2.
AIRA₃ and Muse Spark 1.3 form Meta's twin release matrix at the start of September: Muse Spark 1.3 targets consumer and enterprise API users while AIRA₃ is positioned for internal research-institution use. Together they illustrate Meta's parallel investment in frontier models and research agents.