This model was retired on 30 October 2026 and no longer accepts API requests. Recommended replacement: Muse Glimmer 30B.
Nemotron-3-Super-120B
Retired
Agentic reasoning at scale — 120B total parameters, 12B active, built for coding, planning, tool use, and long-context tasks.
Total Parameters
120B
12B Active
Context Window
256K
Tokens
Quantization
NVFP4
Optimized
Architecture
Hybrid
Mamba-Transformer
Agentic Reasoning at Scale
NVIDIA Nemotron 3 Super 120B A12B NVFP4 is an open hybrid Mamba-Transformer LatentMoE model with 120 billion total parameters and 12 billion active parameters, built for agentic reasoning workloads such as coding, planning, tool use, and long-context tasks. It sits in the same capability tier as Qwen3.5-122B non-reasoning and ahead of GPT-OSS-120B, while also delivering higher throughput.
Hybrid Mamba-Transformer — 12B Active
Built for agentic workloads
Agentic Reasoning
Multi-step reasoning workflows with planning, self-correction, and autonomous decision-making for complex agentic tasks.
Coding & Tool Use
Advanced code generation, debugging, and tool orchestration with native function calling support for engineering workflows.
Long-Context Tasks
Process and reason over massive documents, codebases, and knowledge bases with 256K token context window support.
Planning & Orchestration
Decompose complex goals into executable plans, coordinate multi-agent systems, and orchestrate sophisticated processing pipelines.
Agentic Intelligence
Artificial Analysis Intelligence Index v4.3 scores for the 120B weight class.
Intelligence Index
Better than 72% of models
GPQA Diamond
Better than 86% of models
τ²-Bench Telecom
Better than 82% of models
| Category | Benchmark | Score | Description |
|---|---|---|---|
| Reasoning | GPQA Diamond | 83% | Graduate-level scientific reasoning |
| Reasoning | Humanity's Last Exam | 19% | Humanity's Last Exam |
| Reasoning | τ²-Bench Telecom | 80% | AI agents in dual-control scenarios |
| Reasoning | AA-LCR | 56% | Long context reasoning evaluation |
| Reasoning | IFBench | 71% | Instruction-following accuracy |
| Reasoning | GDPval-AA | 25% | Agentic performance on real-world work tasks |
| Coding | SciCode | 38% | Python for scientific computing |
| Coding | Terminal-Bench Hard | 29% | Agentic coding & terminal use |
| Knowledge | AA-Omniscience Accuracy | 23% | Proportion of correctly answered questions |
| Knowledge | AA-Omniscience Non-Hallucination | 15% | Confidently answered questions that are correct |
Metrics sourced from Artificial Analysis. Evaluated on BF16 weights in regular (highest-effort) reasoning mode.