Meta's AIRA-3 Wins Kaggle Gold, Beating Human AI Teams

Meta says its autonomous AI research system, AIRA-3, placed 8th out of roughly 4,000 teams in a live Kaggle competition run by NVIDIA, outperforming human competitors using the same tools and information.

The competition asked entrants to fine-tune a 30-billion-parameter Nemotron model to reason better, with every team graded externally against the same private test set. Meta says AIRA-3's highest-scoring entry combined GPT-5.5 running inside OpenCode with Claude 4.8 running inside Claude Code.

Rather than routing work through one central controller, Meta says AIRA-3 runs many long-running agents in isolated environments that coordinate asynchronously through two shared substrates: a forum for posting hypotheses and findings, and a shared filesystem for solution artifacts. Agents build on each other's discoveries as the system runs, rather than executing a fixed plan.

Meta says it also tested other model combinations after the competition closed, grading them against the same private test set. A Muse Spark 1.2 pairing reached gold-medal level performance as well, while a Muse Spark 1.1 pairing and a GLM 5.2 pairing each reached silver-medal level, according to Meta.

Meta frames the result as a signal that AIRA-3 can improve a specific AI capability, in this case reasoning in a 30B model, at a level comparable to human experts working with the same frontier tools. The claim rests entirely on Meta's own account of the competition and its post-hoc testing; neither NVIDIA nor Kaggle has independently confirmed Meta's description of the results in the material reviewed here.