The strongest US open-weights model at the time is fast and loses little accuracy at NVFP4. The turn-budget method is a reusable way to compare AI agentAn AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.Full definition → latency fairly.
Key takeaways · AI-distilled
Nemotron 3 Ultra has about 550B total and 55B active parameters and scores 47.7 on the Artificial Analysis Intelligence Index with NVFP4 weights, versus 48.2 in BF16, so the recommended quantizationShrinking a model by storing its numbers less precisely — like rounding — so it runs faster and fits on smaller hardware, at a small quality cost.Full definition → costs little.
It leads US open weightsA model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.Full definition → models on the Intelligence and Agentic indexes, but Gemma 4 31B scores about 1 point higher on the Coding Index, and the Chinese-led open frontier (Kimi K2.6 at 53.9) stays ahead.
Artificial Analysis ran Terminal-Bench v2.1 at 10, 20, 50 and 100 turn budgets, stating the budget up front and warning near the limit, because silently cutting agents off penalizes models that did not know the limit was coming.
Measured pre-release on BlackBox AI at over 400 output tokenThe chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.Full definition → per second, Nemotron 3 Ultra was the fastest model at every turn budget and sat on the Pareto frontier for performance versus time per task.
Terms in this piece · Glossary
token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
quantization — Shrinking a model by storing its numbers less precisely — like rounding — so it runs faster and fits on smaller hardware, at a small quality cost.
open weights — A model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.
