NVIDIA Nemotron 3 Super 120B A12B (Q6_K_XL) — 133 GBon NVIDIA RTX A6000 384GB

Code Multilingual Thinking Tool Calls

Q6_K_XL NVIDIA RTX A6000 384GB

Overview

NVIDIA Nemotron 3 Super 120B A12B is a 123.61B parameter moe language model by NVIDIA, with code, multilingual, thinking, tool-calls capabilities. It supports a context window of up to 262,144 tokens.

Nemotron 3 Super 120B A12B is a 123.61-billion-parameter hybrid Mamba-2 Transformer LatentMoE model from NVIDIA, activating 12 billion parameters per token across 22 of 512 routed experts plus 1 shared expert. Trained on over 25 trillion tokens, it targets agentic reasoning, code generation, tool calling, and multilingual conversation in 7 languages. A 256K context window, toggleable thinking mode, and multi-token prediction enable high-throughput inference for complex multi-agent workflows. Its MoE sparsity quantizes well to GGUF for self-hosted deployment on multi-GPU setups.

At Q6_K_XL quantization (high quality tier), the model weighs 109.75 GB. This fits within the 384 GB of VRAM on NVIDIA RTX A6000 384GB, enabling full GPU inference.

Hardware Requirements

Model size	109.75 GB
VRAM available	384 GB
VRAM used	133 GB
System RAM
Min RAM required	0 GB
GPU layers	88 / 88
Context size	262,144
Backend	cuda13
Flash attention	Yes
Reading from disk	Yes

Performance Notes

Deploy

Prerequisites

Ensure your GPU nodes are prepared with the NVIDIA container toolkit:

ansible-playbook prositronic.infra.nvidia_container_toolkit

Command

helmfile --state-values-file <(curl -s https://www.prositronic.eu/values/nemotron-3-super-120b-a12b/q6_k_xl/nvidia-a6000-384gb.yaml) apply

Generated `values.yaml`

/values/nemotron-3-super-120b-a12b/q6_k_xl/nvidia-a6000-384gb.yaml

Loading values…

Install llama.cpp

Install llama.cpp from the official build instructions for your platform and backend.

Download Model

curl -L -o nemotron-3-super-120b-a12b.gguf "https://huggingface.co/unsloth/NVIDIA-Nemotron-3-Super-120B-A12B-GGUF/resolve/main/UD-Q6_K_XL/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q6_K_XL-00001-of-00004.gguf"

Start Server

llama-server \
  -m nemotron-3-super-120b-a12b.gguf \
  --n-gpu-layers 88 \
  --ctx-size 262144 \
  --flash-attn

Verify

curl http://localhost:8080/health

Frequently Asked Questions

How much VRAM does NVIDIA Nemotron 3 Super 120B A12B (Q6_K_XL) need?

The Q6_K_XL quantization of NVIDIA Nemotron 3 Super 120B A12B requires 109.75 GB. All 88 layers fit in the 384 GB of VRAM available on NVIDIA RTX A6000 384GB, enabling full GPU acceleration.

Can I run NVIDIA Nemotron 3 Super 120B A12B on NVIDIA RTX A6000 384GB?

Yes. NVIDIA RTX A6000 384GB provides 384 GB of VRAM, which is enough to run NVIDIA Nemotron 3 Super 120B A12B (Q6_K_XL) with all layers on the GPU for optimal performance.

What is quantization?

Quantization reduces a model's numerical precision from its original floating-point format to a more compact representation. This shrinks the file size and VRAM footprint, making it possible to run large models on consumer hardware. The trade-off is a small reduction in output quality. Q6_K_XL compresses NVIDIA Nemotron 3 Super 120B A12B from its original size down to 109.75 GB.

What quantization should I choose for NVIDIA Nemotron 3 Super 120B A12B?

Q6_K_XL is a high-quality quantization. Higher-quality quants (Q8, Q6) preserve more model accuracy but need more VRAM. Lower quants (Q4, Q3, Q2) reduce VRAM usage at the cost of some quality. Choose based on your available hardware and quality requirements.

What is flash attention and why is it enabled?

Flash attention is a memory-efficient algorithm that speeds up the attention mechanism in transformer models. It reduces VRAM usage during inference by avoiding the materialisation of the full attention matrix. For NVIDIA Nemotron 3 Super 120B A12B on NVIDIA RTX A6000 384GB, flash attention is enabled to maximise context length and throughput within the available 384 GB of VRAM.

What is MoE and how does it affect deployment?

NVIDIA Nemotron 3 Super 120B A12B uses a Mixture-of-Experts (MoE) architecture with 512 experts, of which 22 are active per token. This means only a fraction of the model weights are used for each inference step, allowing MoE models to be larger in total parameter count while remaining efficient at inference time.

Last updated: March 14, 2026