NVIDIA Nemotron 3 Super 120B A12B
NVIDIA · released 2026-03-01 · NVIDIA Open Model License license
NVIDIA Nemotron 3 Super is a 120B-total / 12B-active hybrid Mamba-Transformer MoE reasoning model (March 2026) tuned for throughput and 1M context. Artificial Analysis index 36, ahead of GPT-OSS-120B, behind Qwen3.5-122B.
Key specs
| Type | Local open-weight |
|---|---|
| Parameters | — |
| Architecture | — |
| Context window | — |
| Knowledge cutoff | — |
| Modalities | text |
| Recommended backends | — |
| Minimum viable rig | — |
Benchmark scores
| GPQA Diamond | 79.23% |
|---|---|
| SWE-bench Verified | 60.47% |
| AIME | 90.21% |
| MMLU-Pro | 83.73% |
| BFCL v3 (tool use) | 61.15% |
| Composite score | 6.54 |
| Community rating | No reviews yet |
Strengths & weaknesses
Strengths: Very high throughput (hybrid Mamba-Transformer MoE, 12B active); Strong long context: RULER 91.75 at 1M tokens; Native 1M-token context
Weaknesses: Lags Qwen3.5-122B on most knowledge/agentic benchmarks; Low HLE (18.26); restrictive NVIDIA license