Is NVIDIA AVO available to download or use?

NVIDIA has not released code, weights or an API for AVO. The developer blog points only to the research paper, 'AVO: Agentic Variation Operators for Autonomous Evolutionary Search', on arXiv. Teams who want the approach get the paper's description of the agent loop, persistent memory and supervisor, but there is no NVIDIA implementation to run today.

Does the 100% score cover all of ARC-AGI-3?

No. NVIDIA states the AVO results cover the 25-environment ARC-AGI-3 public set using the official scorecard and RHAE metric, and are not results on the semi-private or fully private competition sets. ARC Prize keeps those held-out sets to guard against overfitting, so the 100.00 RHAE figure is a public-set number only.

How does AVO compare with Claude Opus 5 on its own?

Claude Opus 5 by itself reaches roughly 30% on the ARC-AGI-3 public set, while AVO reaches 100.00 RHAE using the same model family. NVIDIA cautions that its run used a different reasoning setting and a substantially different agent system, and says the comparison should not be read as a controlled ablation.

What does the RHAE score in ARC-AGI-3 measure?

RHAE stands for Relative Human Action Efficiency. The ARC-AGI-3 metric combines task completion with per-level action efficiency measured against first-time human baselines, then aggregates across levels and environments. AVO used 6,624 environment actions to finish 183 levels, about 12% fewer than the 7,542 actions reported for VISTA.

What else did NVIDIA test AVO on besides the benchmark?

NVIDIA ran AVO on GPU kernel optimization for multi-head attention on DGX B200 systems, operating continuously for seven days. AVO explored more than 500 optimization directions and produced 40 committed kernel versions, ending with kernels that outperform cuDNN by up to 3.5% and FlashAttention-4 by up to 10.5% across the evaluated configurations.

NVIDIA · 2026-08-21 · major

NVIDIA AVO — an agent system scores 100% on ARC-AGI-3's public set

NVIDIA AVO scored 100.00 RHAE on the 25-environment public set of ARC-AGI-3, clearing all 183 levels in 6,624 actions. Claude Opus 5 on its own scores about 30% on the same set. The private competition sets were not tested.

Abstract agentic AI graphic from NVIDIA's AVO announcement

NVIDIA's AVO agent cleared every public ARC-AGI-3 environment, a benchmark where the best single model sat near 30%.

Key specs

Arc agi 3 rhae (public set)	100.00
Levels cleared	183 of 183

Quick facts

Maker	NVIDIA
Full name	Agentic Variation Operators
Benchmark tested	ARC-AGI-3 public set, 25 environments
Base model	Claude Opus 5, with tests on GPT-5.6 Sol
Actions used	6,624 (VISTA reported 7,542)
GPU kernel result	Up to 3.5% over cuDNN, 10.5% over FlashAttention-4
Availability	Research paper only — no code or weights released

What is it?

NVIDIA AVO — short for Agentic Variation Operators — cleared all 25 environments of the ARC-AGI-3 public set with a 100.00 RHAE score. ARC-AGI-3 drops an agent into an unfamiliar turn-based world with no instructions, so it has to work out the goal and the rules by playing. NVIDIA says the run used Claude Opus 5 as the underlying model, with extra testing on GPT-5.6 Sol.

How does it work?

The design replaces the fixed mutation and crossover steps of classic evolutionary search with a coding agent that decides what to try next. Persistent memory carries earlier implementations, evaluation results, compiler and profiler output forward, so the agent resumes work instead of rebuilding its context each round. A separate supervisor watches the wider search for stalling or repeated dead ends and can redirect the main agent to another strategy.

Why does it matter?

Long-horizon autonomy is the weak spot in today's agents, and this result points at the harness around a model mattering as much as the model itself. NVIDIA also ran AVO for seven straight days on GPU kernel work for multi-head attention, where it explored more than 500 optimization directions and committed 40 kernel versions. Teams building agents get a concrete design to copy: memory, a supervisor, grounded feedback and recovery.

Who is it for?

agent researchers and systems engineers

Frequently asked questions

Is NVIDIA AVO available to download or use?: NVIDIA has not released code, weights or an API for AVO. The developer blog points only to the research paper, 'AVO: Agentic Variation Operators for Autonomous Evolutionary Search', on arXiv. Teams who want the approach get the paper's description of the agent loop, persistent memory and supervisor, but there is no NVIDIA implementation to run today.
Does the 100% score cover all of ARC-AGI-3?: No. NVIDIA states the AVO results cover the 25-environment ARC-AGI-3 public set using the official scorecard and RHAE metric, and are not results on the semi-private or fully private competition sets. ARC Prize keeps those held-out sets to guard against overfitting, so the 100.00 RHAE figure is a public-set number only.
How does AVO compare with Claude Opus 5 on its own?: Claude Opus 5 by itself reaches roughly 30% on the ARC-AGI-3 public set, while AVO reaches 100.00 RHAE using the same model family. NVIDIA cautions that its run used a different reasoning setting and a substantially different agent system, and says the comparison should not be read as a controlled ablation.
What does the RHAE score in ARC-AGI-3 measure?: RHAE stands for Relative Human Action Efficiency. The ARC-AGI-3 metric combines task completion with per-level action efficiency measured against first-time human baselines, then aggregates across levels and environments. AVO used 6,624 environment actions to finish 183 levels, about 12% fewer than the 7,542 actions reported for VISTA.
What else did NVIDIA test AVO on besides the benchmark?: NVIDIA ran AVO on GPU kernel optimization for multi-head attention on DGX B200 systems, operating continuously for seven days. AVO explored more than 500 optimization directions and produced 40 committed kernel versions, ending with kernels that outperform cuDNN by up to 3.5% and FlashAttention-4 by up to 10.5% across the evaluated configurations.