Model reference · open weights

NVIDIA-Nemotron-3-Super-UD-Q4_K_XL-MTPv2-layers

LLMs meshllm Text gen · MoE 1 build Its own licence terms 5k dl/mo

NVIDIA-Nemotron-3-Super-UD-Q4_K_XL-MTPv2-layers is an open-weight language model from meshllm. NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-MTPv2-layers (BF16) weighs 90.4 GB; the smallest configuration that runs it is 4× RTX 4090 24 GB.

NVIDIA-Nemotron-3-Super-UD-Q4_K_XL-MTPv2-layers is a GGUF layer package by meshllm designed for text generation across a local Mesh LLM cluster. It splits the 120B-A12B model into 89 per-layer artifacts to enable distributed inference on multiple machines. The package uses UD-Q4_K_XL quantization and is licensed under an other licence.

Summary of the meshllm/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-MTPv2-layers model card, 2026-10-01

Hugging Face ↗

What it is

Released by	meshllm
Type	Language models
Task	Text gen · MoE
Runs with	mesh-llm
Based on	meshllm/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-MTPv2-GGUF
Released	2026-09-12
Popularity	5k downloads / month
Weights	90.4 GB (NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-MTPv2-layers (BF16), file size)
Licence	Its own licence terms

What it runs on

Memory and cards for NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-MTPv2-layers (BF16)

Weights 90.4 GB (file size) · runtime overhead from 651 MB on a small card.

How much memory each request adds is not estimated yet for this architecture — only the weights are. They need the cards below at the least, plus room for the context.

Card	The weights alone
RTX 3060 12 GB … H100 80 GB 8 smaller cards	does not fit
RTX PRO 6000 Blackwell 96 GB	tight
DGX Spark (GB10) 128 GB unified	fits
H200 141 GB	fits
B200 180 GB	fits
4× RTX 4090 24 GB split by layers (llama.cpp)	tight
4× RTX 3090 24 GB split by layers (llama.cpp)	tight
4× RTX 5090 32 GB split by layers (llama.cpp)	fits
2× H100 80 GB split by layers (llama.cpp)	fits
2× A100 80 GB split by layers (llama.cpp)	fits
4× L40S 48 GB split by layers (llama.cpp)	fits
2× RTX PRO 6000 Blackwell 96 GB split by layers (llama.cpp)	fits

From the model card

What meshllm says about NVIDIA-Nemotron-3-Super-UD-Q4_K_XL-MTPv2-layers

Read the model card

GGUF layer package for running NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-MTPv2 across a local Mesh LLM cluster.

This package is derived from meshllm/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-MTPv2-GGUF and keeps the original GGUF distribution split into per-layer artifacts for distributed inference.

Highlights

Run locally	Pool multiple machines	OpenAI-compatible	Package variant
Private inference on your hardware	Split layers across peers	Serve `/v1/chat/completions` locally	`UD-Q4_K_XL` layer package

Model Overview

Property	Value
Source model	meshllm/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-MTPv2-GGUF
Model id	`meshllm/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-MTPv2-GGUF:UD-Q4_K_XL-MTPv2`
Family	NVIDIA
Parameter scale	120B-A12B
Quantization	`UD-Q4_K_XL`
Layer count	89
Activation width	not recorded
Package size	0 B
Source file	`NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-MTPv2-00001-of-00003.gguf`
Package repo	meshllm/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-MTPv2-layers
License	`other` from meshllm/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-MTPv2-GGUF

Recommended Use

Local and private inference with Mesh LLM.
Multi-machine serving when the full GGUF is too large for one host.
OpenAI-compatible chat/completions workflows through Mesh LLM's local API.

For upstream architecture details, chat template guidance, sampling recommendations, license terms, and benchmark notes, see the source model card: meshllm/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-MTPv2-GGUF.

Quickstart

# Run this on each machine that should contribute memory/compute.
mesh-llm serve --model "meshllm/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-MTPv2-layers" --split

# Check the mesh and discover the OpenAI-compatible model name.
curl -s http://localhost:3131/api/status
curl -s http://localhost:3131/v1/models

# Send an OpenAI-compatible chat request.
curl -s http://localhost:3131/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "meshllm/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-MTPv2-GGUF:UD-Q4_K_XL-MTPv2",
    "messages": [{"role": "user", "content": "Write a tiny hello-world function in Rust."}],
    "max_tokens": 128
  }'

Package Variant

Property	Value
Format	`gguf`
Canonical source ref	`meshllm/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-MTPv2-GGUF@360a2016f3b898925d7dc460d717852ba942948e/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-MTPv2-00001-of-00003.gguf`
Source revision	`360a2016f3b898925d7dc460d717852ba942948e`
Source SHA-256	`e0868d034ab4d3b56a74cfa9786894719c147f851149fe256b79113fe4245bc2`
Skippy ABI	`not recorded`
Package manifest SHA-256	`29c3756b4cc31800322a45dcf691303ec37096930ee69a1f395616b56b2cd69d`

What Is Included

Artifact	Path	Contents	SHA-256
Manifest	`model-package.json`	Package schema, source identity, checksums	`29c3756b4cc31800322a45dcf691303ec37096930ee69a1f395616b56b2cd69d`

Validation

Generated by the Mesh LLM HF Jobs splitter from mesh-llm ref f932c4d1dc12b3e3a670d5f470cedd5cdcc5db39. Each artifact is checksummed as it is written, uploaded to this repository, and removed from the job workspace before the next artifact is produced.

skippy-model-package write-package "/hf-cache/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-MTPv2-00001-of-00003.gguf" --out-dir "/tmp/meshllm-layer-job-meshllm_NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-MTPv2-layers-1/package"