NVIDIANemotronNemotronH_Omni_Reasoning_V3

NVIDIA Nemotron 3.5 Super VL 120B A12B BF16 — Hardware Requirements & GPU Compatibility

VisionFunctions

Nemotron 3.5 Super VL is NVIDIA's vision-language model for reasoning, coding, tool use, document understanding, and video understanding, with about 120 billion total parameters and 12 billion active per token. It combines Mamba-2 state-space layers, mixture-of-experts layers, and attention with a C-RADIO vision encoder, and includes multi-token prediction for speculative decoding. This is the full-precision BF16 release, meant for customization and evaluation; the card lists 4x H100 as the deployment target. Even quantized, it needs a multi-GPU workstation or a unified-memory machine with well over 64 GB. The model accepts text, images, and video and outputs text. It supports context lengths up to 1 million tokens, and reasoning is configurable: off, medium, or full, set through the chat template. Languages include English, Spanish, French, German, Italian, and Japanese. It uses the OpenMDW 1.1 license, which NVIDIA describes as allowing commercial and non-commercial use. An NVFP4 version exists for more efficient serving, including on DGX Spark.

124.5B params · MoE12B activeOtherOct 20262.2K downloads65 likes12.4K quant downloads

Your hardware

Verdicts below are for the reference RTX 4090. Set yours to personalise the page.

Detection reads your browser's GPU name with no data sent anywhere.

NVIDIA Nemotron 3.5 Super VL 120B A12B BF16 runs on a 96 GB+ Mac or a workstation GPU

Best local pick → Mac Studio M2 Max (96 GB) · Q4_K_M · 87.7 GB · ~26 tok/s

Macs from 64 GB (~44 tok/s)

Other

Not set up for it?

Reference GPU · RTX 4090 24 GB

Estimates, ±20% · how we estimate

Run NVIDIA Nemotron 3.5 Super VL 120B A12B BF16

One recommended way. Every other method, OS and quant is in the step-by-step guide.

Other ways

each opens step-by-step

Run in cloud

Fits on H200 (141 GB) (53 GB headroom) · Q4_K_M

Generation speed: ~193 tok/s; generation speed
Cost per 1M output tokens: $5.16; per 1M output tokens

Deployon RunPodBest optionfrom $3.59/hr

Compare GPUs →

Run NVIDIA Nemotron 3.5 Super VL 120B A12B BF16

87.7 GB VRAM at 16K context · needs offload on the RTX 4090 (63.7 GB short)

Works on macOS, Windows and Linux.

Install Ollama (Linux)
$curl -fsSL https://ollama.com/install.sh | sh
Install Ollama (macOS / Windows)
Download and run the app; it installs the ollama command.
Download Ollama ↗
Run Q4_K_M
$ollama run hf.co/ggml-org/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-GGUF
ggml-org/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-GGUF does not list Q4_K_M; Ollama picks a quant itself (Q4_K_M if present). Check the repo's Files tab.
Raise the context window
Type it inside the chat. For the API server instead, start it with OLLAMA_CONTEXT_LENGTH=16384 ollama serve.
$/set parameter num_ctx 16384
A larger context needs more memory than the weights alone.
Use the API
Ollama serves its API on http://localhost:11434.

Works on macOS, Windows and Linux.

Install LM Studio
Desktop app for macOS (Apple silicon), Windows and Linux.
lmstudio.ai ↗
Download the model
In the app press Ctrl/Cmd + Shift + M and paste https://huggingface.co/ggml-org/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-GGUF, then pick Q4_K_M. Or use the CLI (run LM Studio once first):
$lms get https://huggingface.co/ggml-org/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-GGUF
Open on Hugging Face ↗
Load and chat
Open the model loader, select the model and set Context Length to 16384. Leave GPU offload at maximum if it fits.
Start the local server (optional)
$lms server start

Works on macOS, Windows and Linux.

Install llama.cpp (macOS)
MacPorts, Nix and conda-forge also work.
$brew install llama.cpp
Install options ↗
Install llama.cpp (Windows)
$winget install llama.cpp
Install llama.cpp (Linux)
Download the prebuilt llama-*-bin-ubuntu-* archive for your GPU backend (CPU, Vulkan) from the releases page, unpack it and run the binaries from that folder. Homebrew (brew install llama.cpp) and conda-forge also work on Linux.
llama.cpp releases ↗
Start the server
Downloads the model on first run. The chat UI and an OpenAI-compatible API are at http://localhost:8080.
$llama-server -hf ggml-org/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-GGUF -c 16384 -ngl 99 --port 8080
ggml-org/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-GGUF does not list Q4_K_M; llama.cpp falls back to Q4_K_M or the first file.

Works on macOS, Windows and Linux.

Option A: Docker Model Runner
Docker Desktop: Settings > AI > Enable Docker Model Runner. Docker Engine on Linux: install the docker-model-plugin package.
$docker model run hf.co/ggml-org/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-GGUF
Model Runner docs ↗
The quant tag after the repo is how Hugging Face cards write it; if your Docker version rejects it, drop the tag and pick the quant by repo.
Option B: llama.cpp server image (NVIDIA)
Needs the NVIDIA Container Toolkit. Models are cached in the llama-cache volume. Open http://localhost:8080.
$docker run --gpus all -p 8080:8080 -v llama-cache:/root/.cache ghcr.io/ggml-org/llama.cpp:server-cuda -hf ggml-org/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-GGUF -c 16384 -ngl 99 --host 0.0.0.0 --port 8080
Image tags (cuda13, vulkan, rocm) ↗
Cache path inside the image is not in the docs; remove the -v flag if the volume complains.

Linux only.

Install vLLM
Linux only; on Windows run it inside WSL2. NVIDIA GPU shown; AMD ROCm, Intel and CPU builds have their own install pages.
$uv venv --python 3.12 --seed
$source .venv/bin/activate
$uv pip install vllm --torch-backend=auto
vLLM installation ↗
Serve the model
OpenAI-compatible API at http://localhost:8000/v1.
$vllm serve nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16 --max-model-len 16384
Runs the original ~274 GB weights, not the Q4_K_M GGUF.

macOS only.

MLX supports only macOS, so follow these steps for that setup.

Install mlx-lm
$pip install mlx-lm
Chat in the terminal
$mlx_lm.chat --model Duckets/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-MLX-4bit
Downloads the model on first run. The MLX repo has its own quantisation; it may not match the quant selected above.
Or start a local server
Listens on http://localhost:8080.
$mlx_lm.server --model Duckets/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-MLX-4bit

Works on macOS, Windows and Linux.

Install Jan
jan.ai ↗
Find the model
Open the Hub tab, paste this model ID in Search, and choose Q4_K_M if offered.
$ggml-org/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-GGUF
Use the local API (optional)
Settings > Local API Server > Start Server. It listens on http://127.0.0.1:1337.

Rent GPUs hourly to run Q4_K_M right there. Pricing comes from partner marketplaces and varies.

Available GPU rentals for Q4_K_M, prioritized by value
GPU	Speed	$/1M tokens	Rent
H200 SXM (141 GB)CheapestCheapest53 GB free	~193 tok/s	$5.16	Deploy on RunPodfrom $3.59/hr (opens in new tab)	141 GB · 53 GB free · ~193 tok/s · $5.16 / 1M tokens
H200 NVL (141 GB)53 GB free	~193 tok/s	$5.76	Deploy on Vast.aifrom $4.00/hr (opens in new tab)	141 GB · 53 GB free · ~193 tok/s · $5.76 / 1M tokens
B200 (192 GB)104 GB free	~240 tok/s	$6.92	Deploy on RunPodfrom $5.98/hr (opens in new tab)	192 GB · 104 GB free · ~240 tok/s · $6.92 / 1M tokens
B300 (288 GB)200 GB free	~240 tok/s	$8.03	Deploy on RunPodfrom $6.94/hr (opens in new tab)	288 GB · 200 GB free · ~240 tok/s · $8.03 / 1M tokens
B200 (192 GB)104 GB free	~240 tok/s	$9.96	Deploy on Vast.aifrom $8.61/hr (opens in new tab)	192 GB · 104 GB free · ~240 tok/s · $9.96 / 1M tokens

Compare GPUs →

How Much VRAM Does NVIDIA Nemotron 3.5 Super VL 120B A12B BF16 Need?

Select a quantization level; the listed hardware adheres to it.

VRAM	+32K	+128K	File	Quality	On RTX 4090 (reference)
37.6 GB	49.5 GB	87.5 GB	51.1 GB	2-bit · i-quant extreme compression — Importance-weighted 2-bit, extreme compression — significant quality loss	Does not fit: needs 43.2 GB
46.2 GB	58.1 GB	96 GB	55.9 GB	2-bit · i-quant medium — Importance-weighted 2-bit, medium	Does not fit: needs 51.7 GB
53.0 GB	64.9 GB	102.9 GB	59.8 GB	3-bit · i-quant — Importance-weighted 3-bit	Does not fit: needs 58.6 GB
58.2 GB	70 GB	108 GB	~52.9 GB	2-bit · K-quant — 2-bit quantization with K-quant improvements	Does not fit: needs 63.7 GB
61.6 GB	73.5 GB	111.4 GB	70 GB	3-bit · i-quant medium — Importance-weighted 3-bit, medium	Does not fit: needs 67.1 GB
66.7 GB	78.6 GB	116.5 GB	~60.7 GB	3-bit medium — 3-bit medium quantization	Does not fit: needs 72.3 GB
68.5 GB	80.3 GB	118.3 GB	~62.2 GB	4-bit · legacy — 4-bit legacy quantization	Does not fit: needs 74 GB
77.0 GB	88.9 GB	126.8 GB	86.9 GB	4-bit small — 4-bit small quantization	Does not fit: needs 82.5 GB
77.0 GB	88.9 GB	126.8 GB	74.8 GB	4-bit · i-quant non-linear — Importance-weighted 4-bit, non-linear	Does not fit: needs 82.5 GB
82.1 GB	94 GB	131.9 GB	~74.7 GB	4-bit medium — 4-bit medium quantization — most popular sweet spot	Does not fit: needs 87.7 GB
94.1 GB	106 GB	143.9 GB	110.6 GB	5-bit small — 5-bit small quantization	Does not fit: needs 99.7 GB
97.5 GB	109.4 GB	147.3 GB	~88.7 GB	5-bit medium — 5-bit medium quantization — good quality/size tradeoff	Does not fit: needs 103.1 GB
112.9 GB	124.8 GB	162.7 GB	120 GB	6-bit, very good — 6-bit quantization, very good quality	Does not fit: needs 118.5 GB
136.9 GB	148.8 GB	186.7 GB	~124.5 GB	8-bit, near-lossless — 8-bit quantization, near-lossless	Does not fit: needs 142.4 GB
273.8 GB	285.7 GB	323.6 GB	~248.9 GB	16-bit · bfloat — Brain floating point 16 — preferred for training	Does not fit: needs 279.3 GB

est.calculated, no published file yet

Selected Q4_K_M

4-bit medium quantization

87.7GBat 16K context

Weights: 81.1 GB
KV cache: 6.3 GB
Overhead: 0.3 GB

Doesn't fit RTX 4090 (reference) · needs 63.7 GB more

≥30 tok/s needs ≥396 GB/s · M3 Ultra ~43, M2 Ultra ~42

Or rent H200 from $3.59/hr.

Which Hardware Can Run NVIDIA Nemotron 3.5 Super VL 120B A12B BF16?

Q4_K_M · 82.1 GB

NVIDIA Nemotron 3.5 Super VL 120B A12B BF16 (Q4_K_M) requires 82.1 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 107+ GB is recommended. No consumer GPU has enough memory. 32 devices with unified memory can run NVIDIA Nemotron 3.5 Super VL 120B A12B BF16, including NVIDIA DGX H100, NVIDIA DGX A100 640GB, Mac Studio M2 Max (96 GB).

By VRAM tier · typical GPUs and Macs at Q4_K_M

Typical hardwareQuantVRAMtok/sContext that fits

8–16 GB · RTX 4060, RTX 3060 12GB, RTX 5060 Ti 16GBDoesn't fit — smallest quant needs 43.2 GB8 GB tier →12 GB tier →16 GB tier →

24 GB · RTX 4090referenceDoesn't fit — smallest quant needs 43.2 GB24 GB tier →

Mac 32 GB · Mac Mini M4Doesn't fit — smallest quant needs 43.2 GBMacs →

Mac 64 GB · Mac Studio M4 MaxIQ3_XXS · 58.6 GB · ~44 tok/s · 16K58.6~4416KMacs →

How good is NVIDIA Nemotron 3.5 Super VL 120B A12B BF16?

llmrun Score blends public benchmark boards into one number; 100 = the average open model.

Not enough data yet for an llmrun Score.

A Score needs results on enough independent boards. How the Composite Score is built

Benchmarks

LMArena Text
LMArena · Oct 2026Arena
1415.9
#39/187 open
Arena

NVIDIA Nemotron 3.5 Super VL 120B A12B BF16 benchmark results

Alternatives to NVIDIA Nemotron 3.5 Super VL 120B A12B BF16

Nemotron family · 42 models →

Similar models

6tok/s on RTX 4090

Model	Score	VRAM	tok/s	Why it is similar
Qwen3.5 122B A10BAlibaba	109.4	75.5 GB	—needs > 24 GB	same size
GPT OSS 120BOpenAI	97.6	70.8 GB	—needs > 24 GB	same size
Ling 3.0 Flash VLInclusion AI	—	76 GB	—needs > 24 GB	same size
Leanstral 1.5 119B A6BMistral AI	—	84 GB	—needs > 24 GB	same size
Laguna S 2.1poolside	—	72.5 GB	—needs > 24 GB	same size
XORTRON.CriminalComputing.LARGE.2026.3darkc0de	—	79.8 GB	—needs > 24 GB	same size

Fine-tunes & variants(1)community builds of NVIDIA Nemotron 3.5 Super VL 120B A12B BF16

NVIDIA Nemotron 3.5 Bolt 74B A9B BF16

Specifications

Publisher: NVIDIA
Family: Nemotron
Parameters: 124.5B
Active parameters: 12B per token
Architecture: NemotronH_Omni_Reasoning_V3
Release Date: 2026-10-12
License: Other

Resources

Frequently Asked Questions

How much VRAM does NVIDIA Nemotron 3.5 Super VL 120B A12B BF16 need?

NVIDIA Nemotron 3.5 Super VL 120B A12B BF16 requires 82.1 GB of VRAM at Q4_K_M, or 273.8 GB at BF16.

VRAM = Weights + KV Cache + Overhead

Weights = 124.5B × 4.8 bits ÷ 8 = 74.7 GB

KV Cache + Overhead ≈ 7.4 GB (at 2K context + ~0.3 GB framework)

Fit ratings and hardware model lists check this model with room for a 16K-token context, which needs a little more memory.

VRAM usage by quantization

Q4_K_M

82.1 GB

Learn more about VRAM estimation →

Can NVIDIA GeForce RTX 5090 run NVIDIA Nemotron 3.5 Super VL 120B A12B BF16?

No — NVIDIA Nemotron 3.5 Super VL 120B A12B BF16 requires at least 37.6 GB at IQ2_XXS, which exceeds the NVIDIA GeForce RTX 5090's 32 GB of VRAM.

What's the best quantization for NVIDIA Nemotron 3.5 Super VL 120B A12B BF16?

For NVIDIA Nemotron 3.5 Super VL 120B A12B BF16, Q4_K_M (82.1 GB) offers the best balance of quality and VRAM usage. Q5_K_S (94.1 GB) provides better quality if you have the VRAM. The smallest option is IQ2_XXS at 37.6 GB.

VRAM requirement by quantization

IQ2_XXS

37.6 GB~53%

IQ3_M

61.6 GB~78%

Q4_K_S

77.0 GB~88%

Q4_K_M ★

82.1 GB~89%

Q5_K_M

97.5 GB~92%

BF16

273.8 GB~100%

★ Recommended — best balance of quality and VRAM usage.

Learn more about quantization →

Can I run NVIDIA Nemotron 3.5 Super VL 120B A12B BF16 on a Mac?

Yes, but only at lower quantizations. The smallest Mac that can run NVIDIA Nemotron 3.5 Super VL 120B A12B BF16 is Mac Mini M4 Pro (48 GB) at IQ2_XXS; 56 of the 156 Macs we list can run it at some quantization. For Q4_K_M (82.1 GB) you need a Mac with more unified memory.

Can I run NVIDIA Nemotron 3.5 Super VL 120B A12B BF16 locally?

Yes — NVIDIA Nemotron 3.5 Super VL 120B A12B BF16 can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.

How fast is NVIDIA Nemotron 3.5 Super VL 120B A12B BF16?

At Q4_K_M, NVIDIA Nemotron 3.5 Super VL 120B A12B BF16 can reach ~89 tok/s on AMD Instinct MI350X. Speed depends mainly on GPU memory bandwidth. Real-world results typically within ±20%.

tok/s = 1000 ÷ (active GB ÷ (bandwidth GB/s × efficiency) × 1000 + layers × routing ms)

Mixture-of-Experts: just the active experts get read for each token, along with a fixed routing cost per layer.

Example: NVIDIA B200 → 7.9 GB active ÷ (8000 × 0.65) = 1.52 ms, plus 48 layers × 0.055 ms = 2.64 ms, so 1000 ÷ 4.16 ms = ~240 tok/s

Estimated speed at Q4_K_M (82.1 GB)

NVIDIA B200

~240 tok/s

NVIDIA B300

~240 tok/s

NVIDIA GH200 Grace Hopper Superchip

~195 tok/s

Real-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.

Learn more about tok/s estimation →

What's the download size of NVIDIA Nemotron 3.5 Super VL 120B A12B BF16?

At Q4_K_M, the download is about 74.67 GB. The full-precision BF16 version is 248.91 GB. The smallest option (IQ2_XXS) is 34.22 GB.

Which GPUs can run NVIDIA Nemotron 3.5 Super VL 120B A12B BF16?

No single consumer GPU has enough VRAM to run NVIDIA Nemotron 3.5 Super VL 120B A12B BF16 at Q4_K_M (82.1 GB). Multi-GPU or professional hardware is required.

Which devices can run NVIDIA Nemotron 3.5 Super VL 120B A12B BF16?

32 devices with unified memory can run NVIDIA Nemotron 3.5 Super VL 120B A12B BF16 at Q4_K_M (82.1 GB), including ASUS Ascent GX10, Asus ROG Flow Z13 (2025, Ryzen AI Max+ 395, 128 GB), Beelink GTR9 Pro (Ryzen AI Max+ 395, 128 GB), Framework Desktop (Ryzen AI Max+ 395, 128 GB). Apple Silicon Macs use unified memory shared between CPU and GPU, making them well-suited for local LLM inference.