Nemotron-3 Super (120B-A12B)
NVIDIA's flagship Nemotron-3 Super — 120B-parameter hybrid Mamba/Transformer Mixture-of-Experts (~12B active per token). Native function-calling, 128K context. Frontier-tier on-device reasoning for workstations and Mac Studio-class hardware — needs ~100GB of usable memory for Q4_K_M. MLX path is held until mlx-community publishes a build that doesn't require trust_remote_code; for now this is llama-only.
At a glance
- Parameters
- 120B
- Quantization
- Q4_K_M
- Context window
- 131K tokens
- Approx. size
- 86.1 GB
- Engines
- llama.cpp
- License
- NVIDIA Open Model License (custom-restricted)
- Version
- v1.0.0 , released 2026-05-17
- Family
- nemotron
Tuning
Defaults a gezel client applies out of the box, from real evaluation runs against this exact quantization.
- sampling.temperature
1- sampling.topP
0.95- reasoning.thinkingBudget
4096- reasoning.enableThinking
true- profiles
- thinking-general, thinking-coding, thinking-precise, instruct, creative
Model behaviors
Client-side behavior modules the catalog enables for this model (reasoning-tag handling, fabrication detection, prompt shaping).
- reasoning.strip-think-tags
- prompt.private-reasoning-guidance
- fabrication.detect-past-tense-no-tools
- turn.preamble-folding
Sources
- llama.cpp (GGUF)
- lmstudio-community/NVIDIA-Nemotron-3-Super-120B-A12B-GGUF · Q4_K_M
- Upstream
- https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16
- Manifest
- View on GitHub
Tags
- nvidia
- reasoning
- moe
- hybrid
- tools
- large