Model Details
Nemotron 3 Super 120B A12B
Quick answer: NVIDIA Nemotron 3 Super 120B A12B is NVIDIA's June 2025 mid-tier MoE model scoring 81.2% LiveCodeBench, 83.7% MMLU-Pro, 73.9% Arena Hard v2, and 72.6% IFBench. NVIDIA Open Model License with 120B total / 12B active parameters.
At a Glance
Where Nemotron 3 Super 120B leads
- 81.2% LiveCodeBench — exceptional coding performance; above frontier closed models from the same period
- 83.7% MMLU-Pro — strong academic knowledge
- 73.9% Arena Hard v2 — competitive instruction quality
- 12B active params: efficient MoE inference
Where it lags
- NVIDIA Open Model License — has commercial use restrictions (not fully permissive like Apache 2.0)
- Below Nemotron 3 Ultra 550B on GPQA/HLE/BrowseComp metrics
- June 2025: pre-dates newer Kimi K2.6 / Qwen3.5 models
Best for: NVIDIA NIM coding pipelines; LiveCodeBench-heavy use cases; NVIDIA AI infrastructure with 120B/12B MoE efficiency.
What Nemotron 3 Super 120B A12B Is
Nemotron 3 Super 120B A12B is NVIDIA's June 2025 mid-tier MoE model — the "Super" tier below the "Ultra" (550B) in NVIDIA's Nemotron 3 lineup. With 120B total and 12B active params, it is more efficient than the Ultra for inference while still delivering strong benchmark scores.
The 81.2% LiveCodeBench score is notable — exceeding many proprietary and open models from the same period on this real-world coding benchmark. This makes it a strong choice for coding-agent deployments within NVIDIA infrastructure.
Specifications
| Field | Value |
|---|---|
| Organization | NVIDIA |
| License | NVIDIA Open Model License |
| HuggingFace | nvidia/Nemotron-3-Super-120B-A12B |
| Release date | June 2025 |
| Parameters | 120B total / 12B active (MoE) |
| Modality | Text only |
Pricing
Available under NVIDIA Open Model License via NVIDIA NIM and cloud providers.
Public Benchmark Scores
| Benchmark | Score | Source | Date |
|---|---|---|---|
| LiveCodeBench | 81.2% | Benchgen evaluation | 2025-06 |
| MMLU-Pro | 83.7% | Benchgen evaluation | 2025-06 |
| Arena Hard v2 | 73.9% | Benchgen evaluation | 2025-06 |
| IFBench | 72.6% | Benchgen evaluation | 2025-06 |
Nemotron 3 Super 120B vs Alternatives
| Model | LiveCodeBench | MMLU-Pro | Active Params | License |
|---|---|---|---|---|
| Nemotron 3 Super 120B | 81.2% | 83.7% | 12B | NVIDIA Open |
| Nemotron 3 Ultra 550B | — | 86.8% | 55B | NVIDIA Open |
| Llama 3.1 Nemotron Ultra 253B | 66.3% | — | 253B | Llama 3.1 |
| Qwen3 30B A3B | 62.6% | — | 3B | Apache 2.0 |
Nemotron 3 Super 120B leads on LiveCodeBench (81.2%) among NVIDIA models and open models. For Apache 2.0 alternatives: Qwen3 30B A3B at 3B active params is more efficient but scores 62.6% LiveCodeBench.
Related Models & Benchmarks
- Compare: Nemotron 3 Ultra 550B A55B · Qwen3 30B A3B · Llama 3.1 Nemotron Ultra 253B
- Benchmarks: LiveCodeBench · MMLU-Pro · Arena Hard
Frequently Asked Questions
What is Nemotron 3 Super 120B A12B?
NVIDIA's June 2025 mid-tier MoE model scoring 81.2% LiveCodeBench, 83.7% MMLU-Pro, 73.9% Arena Hard v2. NVIDIA Open Model License with 120B/12B active params.What is NVIDIA Open Model License?
NVIDIA's custom open model license allows use and distribution but may have commercial restrictions depending on usage scale. It is less permissive than Apache 2.0 or MIT. Review the full license terms at NVIDIA's model page before commercial deployment.Specs from NVIDIA's Nemotron 3 Super 120B A12B release (June 2025) and Benchgen evaluations. Last updated 2026-07-24.
Benchmark Leaderboards
This model isn’t on any benchmark leaderboard yet.