Meta says AIRA₃, the newest version of its autonomous AI research system, was entered in June into a public Kaggle contest that Nvidia was running, and placed 8th among some 4,000 teams for a Gold medal. The task asked every entrant to fine-tune Nvidia’s 30 billion parameter Nemotron model so it reasoned more effectively, with all teams starting from the same materials and scored by an external, private test set. Meta says AIRA₃ beat human teams working with the same frontier tools, though it has not published the full leaderboard or named who finished ahead of or behind it.
That detail is the more interesting part of the announcement. Most agentic capability claims from AI labs come from tests the lab itself designed and graded. A public Kaggle leaderboard, judged by an outside private test set and populated by roughly 4,000 independent entries, gives outsiders a shared task to check a claim against, which is a genuinely stronger form of evidence than another benchmark table published by the company that built the system.
It is still, however, Meta’s own account of what happened. The company has not disclosed AIRA₃’s compute budget, run time, or how its entry compares against the specific human teams it says it outperformed, and no outside party has independently confirmed the placement beyond what Kaggle’s leaderboard already shows publicly.
AIRA₃’s architecture skips a single controlling model. It operates numerous long-running agents at once, each an isolated pairing of a model and a coding tool, and coordinates them through two shared resources instead of top-down direction: a forum where agents post hypotheses and findings, and a shared filesystem holding solution files that later agents can build from.
The 8th-place entry ran GPT 5.5 inside OpenCode alongside Claude 4.8 inside ClaudeCode. After the leaderboard closed, Meta says it tried other pairings against the identical private test set: Muse Spark 1.2 paired with MuseCode also cleared the Gold threshold, and both Muse Spark 1.1 paired with OpenCode and GLM 5.2 paired with OpenCode landed at Silver level. Those later runs carried no competitive stakes since the contest had already ended.
Meta also cites two results outside the Nemotron task. In an internal test, the company says AIRA₃ cut latency by 27 percent on GPU kernels running in its own production systems, a figure Meta measured itself rather than one graded by an outside party. Separately, the system reportedly earned a Gold-level result in a different Kaggle contest that asked entrants to render ancient Akkadian writing, pressed into clay tablets roughly 4,000 years ago, into English.
Meta frames the project as evidence that letting many agents compound knowledge over time, rather than running one agent harder, is the path toward research systems that improve themselves. The company acknowledges the effort is early and says harder problems remain ahead.
Operators building or buying autonomous research agents should treat the Nemotron result as the more credible data point here (it was graded outside Meta’s walls) and treat the GPU kernel figure as an internal claim pending its own outside test. Anyone evaluating AIRA₃-style architectures should ask a vendor to point to a public leaderboard result before accepting an internal benchmark as proof of agentic research capability.
Meta AI described AIRA₃’s Kaggle results in a thread posted to its own account on X.
