#12NVIDIA· released 2026-06-04

NVIDIA: Nemotron 3 Ultra.

NVIDIA Nemotron 3 Ultra is an open frontier-reasoning and orchestration model from NVIDIA, with 55B active parameters out of 550B total (MoE). Built on a hybrid Transformer-Mamba mixture-of-experts architecture, it...

View on OpenRouter ↗
Context
262K
Input $/1M
$0.60
Output $/1M
$2.40
Modality
text

Benchmarks

How these are aggregated →
Arena Elo
1582
LMArena — human blind votes
AA Index
59.0
Artificial Analysis composite
LiveBench
80.3
Contamination-resistant
GPQA
85.7%
Grad-level science reasoning
MMLU-Pro
85.3%
Broad knowledge
SWE-bench
70.8%
Real GitHub bugs (Verified)
Terminal-Bench
58.1%
Agentic shell tasks
ARC-AGI
30.7%
Abstract reasoning
Aider Polyglot
79.6%
Multi-language code editing
Speed
68 tok/s
Median output speed

Profile

NVIDIA: Nemotron 3 Ultra is NVIDIA's entry at position #12 on the current aggregated leaderboard. Its strongest signals are graduate-level science reasoning, real-world software engineering, unattended agentic work. Every figure below is aggregated from the public sources listed on the homepage and refreshed on each pipeline run.

The score spread is comparatively complete for this model, but the standing advice still applies: treat gaps of a few points as ties and validate on your own workload before committing.

How this model's position is computed is documented in the methodology; what each benchmark actually measures is in the benchmark guide. See a figure that disagrees with its source? Report it — accepted corrections apply on the next refresh.

Nearby on the leaderboard