NVIDIAopen weightsCurrentReasoning
Nemotron 3.5 Super VL
120B-A12B hybrid Mamba-Transformer MoE with vision (NVIDIA Nemotron 3.5 Super VL 120B-A12B), distilled from Nemotron 3 Ultra; released in BF16 and NVFP4. No NVIDIA list price (NIM / self-host). NVIDIA reports Terminal-Bench 2.1 60.7 (Terminus 2 harness; 71.9 with parallel execution + self-verification), SWE-bench Pro 51.8.
Oct 15, 2026
—
—
1M
Benchmark scores
| Benchmark | Score | Δ vs prior | Rank (current) | Configuration | Source |
|---|---|---|---|---|---|
SWE-bench Verified | 71.1% | #6 of 8 | NVIDIA technical report, OpenCode harness · lab-reported | research.nvidia.com ↗ | |
GPQA Diamond | 89.5% | #11 of 22 | NVIDIA technical report (89.52), no tools · lab-reported | research.nvidia.com ↗ | |
Humanity's Last Exam | 30.7% | #29 of 36 | Artificial Analysis run, no tools (2,158 text-only questions); NVIDIA reports 31.8 no tools (text), 46.0 with tools · independent run | artificialanalysis.ai ↗ | |
ARC-AGI-2 | 1.9% | #14 of 14 | ARC Prize semi-private, medium reasoning (snapshot 2026-10-28) · leaderboard | benchlm.ai ↗ | |
τ²-bench (Telecom) | 73.2% | #5 of 8 | NVIDIA technical report, Telecom domain (73.19) · lab-reported | research.nvidia.com ↗ | |
Artificial Analysis Intelligence Index | 14 | #30 of 33 | Artificial Analysis Intelligence Index v5 · independent run | artificialanalysis.ai ↗ |