NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16: 80GB+ VRAM, GGUF Builds

NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16: 80GB+ VRAM, GGUF Builds

At a Glance

Item Value
Repository nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16
Family guide NVIDIA-Nemotron-3.5-Super-VL guide (2 articles)
Publisher guide NVIDIA: models and licenses
Published 2026-10-13
License openmdw-1.1
Formats safetensors
Source type Primary source (the publisher itself)

Values determined by this site’s code when the information was collected. Dates are JST.

Overview

NVIDIA Corporation has released “NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16", the full-precision version of its vision-language model supporting inference, coding, tool use, document understanding, and video comprehension. This BF16 model is provided for developers aiming for customization, domain adaptation, and full-precision evaluation. According to the model card, it adopts a hybrid architecture combining Mamba-2, Mixture-of-Experts (MoE), and attention layers with the C-RADIO vision encoder and a vision-language connector.

Specifications

  • Architecture: Mamba-2–Transformer hybrid Mixture of Experts (MoE) with Multi-Token Prediction (MTP)
  • Parameters: 120B total / 12B active
  • Context Length: Up to 1M tokens

Performance

Benchmark comparison tables from the model card (including other open-weight models) are as follows:

→ Scroll horizontally to see all columns

Benchmark Nemotron 3.5 SuperGA BF16 Qwen3.5-122B-A10B Qwen3.8-Flash-Next DeepSeek-V4-Flash0731 Nemotron 3 Ultra
GPQA-Diamond 89.52 86.7 92.7 89.6 87.0
HLE, text-only 31.79 24.2 39.6 35.8 29.2
MMLU-ProX 76.91 83.2 83.0 85.5 83.1
LiveCodeBench v6 88.22 76.7 94.1 94.6 89.2
SWE-bench Verified 71.10 71.9 78.7 75.4 73.9
SWE-bench Pro 51.78 44.9 60.37 56.91 49.66
Terminal-Bench 2.1 66.70 48.70 85.80 76.80 54.80
OSWorld-Verified 51.14 51.5 55.8 N/A N/A

According to the evaluation results published by the authors, it scores 89.52 on GPQA-Diamond, which measures difficult scientific questions, outperforming some of the comparison models. It also records high figures on LiveCodeBench v6, which handles new competitive programming problems, and Terminal-Bench 2.1, which evaluates agent-based terminal operation. On the other hand, it falls behind other comparison models in general knowledge metrics such as MMLU-ProX, and does not achieve the highest values in every field.

Strengths and Use Cases

The primary feature of this model is its ability to integrate and process three modalities: text, images, and video. In particular, because it supports an extremely long context length of up to 1M (1 million) tokens, it is suited for understanding vast documents spanning thousands of pages, analyzing video content of up to 2 minutes, and performing advanced reasoning while maintaining complex dialogue history.

Specific strengths and expected use cases include:

  • Multimodal Document and Video Understanding: Equipped with the C-RADIO vision encoder, it excels in OCR capabilities to accurately read text and formulas within images, and the ability to grasp events in videos chronologically. It demonstrates high performance in tasks such as document analysis and video summarization.
  • Advanced Coding and Agent Tasks: Having undergone mid-training specialized for software engineering, it is oriented toward solving real GitHub issues and acting as an agent that executes commands in a terminal to complete work. High scores on LiveCodeBench and SWE-bench back up its capability.
  • Native Speculative Decoding: Because it incorporates a Multi-Token Prediction (MTP) head, it can perform speculative decoding with the model alone, without preparing an additional draft model. This is expected to improve inference speed.
  • Multilingual Support and Customization: It supports multiple major languages including Japanese. Furthermore, since it is released as a BF16 full-precision version, it is optimal as a base model for fine-tuning specialized in specific business domains or creating proprietary quantized models.

These capabilities are built upon a massive dataset containing over 20 trillion tokens of foundational training data, more than 10 trillion tokens of text data, 1 million to 1 billion images, and 10,000 to 1 million hours of video. In particular, data in specialized domains such as mathematics, science, law, and finance, as well as high-quality code data, have been learned with emphasis.

Hardware Requirements

Estimated requirements (calculated by Local Model Watch) — 124.5B parameters

Your VRAM Quantization File size Est. memory needed
80GB class (A100 / H100) Q3_K_XL 61.2GB 73.5GB

Inference engine support (architecture name matched against each project’s own model registry in its source code, checked 2026-11-17): llama.cpp: registered, vLLM: registered, MLX (mlx-lm): not registered. “Not registered" means the name is absent from that registry today, not that the model cannot run.

Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. File sizes are measured from the converted build unsloth/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-GGUF. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.

Can You Run It Locally?

Runs in Ollama, LM Studio and llama.cpp via a converted build.

The publisher ships safetensors, but unsloth/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-GGUF provides a GGUF build you can use.

License — openmdw-1.1: A custom license from the publisher. Check the original terms directly, including whether commercial use is permitted.

Compiled by this site’s code from the published formats, converted builds we have found, and each engine’s own model registry. “Not found" means we have not seen such a build, not that none exists. License summaries are not legal advice — check the publisher’s original terms before relying on them.

Our Own Measurements

Measurements We Did Not Take

We have not confirmed that this site meets the commercial-use terms of this model’s license (openmdw-1.1). Because this site carries advertising, we did not run the model (no CPU run, answers, quantization comparison, conversion or generation).

Recent Models in the Same Size Class

Models with over 40B parameters that Local Model Watch covered recently, listed by code from our article log for comparison. VRAM tiers are this site’s estimates; licenses are as stated on the model cards.

→ Scroll horizontally to see all columns

Model Parameters Smallest VRAM tier License Our article
deepseek-ai/DeepSeek-V4.1-Flash-1017 763.2B — mit DeepSeek-V4.1-Flash-1017 Vision-Language Model: ~570GB Memory (2026-11-17)
Qwen/Qwen3.8-Flash-Next 180.0B — qwen-community-1.0 Qwen3.8-Flash-Next Multimodal MoE Model: ~402GB Memory (2026-10-04)
ggml-org/GLM-5.3-Flash-GGUF 321.3B — other GLM-5.3-Flash-GGUF Vision-Language Model: ~150GB Memory (2026-10-03)
bartowski/Intern-S2-397B-GGUF 403.4B — apache-2.0 Intern-S2-397B-GGUF Vision-Language Model: ~102GB Memory (2026-09-15)
deepseek-ai/DeepSeek-V4.1-Flash 763.2B — mit DeepSeek-V4.1-Flash 552B Multimodal MoE Model: ~570GB Memory (2026-09-10)

How to Get It

This model is available in safetensors format from the Hugging Face repository. The “OpenMDW-1.1" license applies, allowing use for both commercial and non-commercial purposes. At this time, gated settings requiring individual license agreement upon acquisition are not applied.

Recommended execution environments include vLLM (nightly version or v0.32.0 and later) and SGLang (v0.5.20 and later). Operation in BF16 full precision requires high computational resources, and the official documentation exemplifies configurations using 4 H100 GPUs or 2 GB200 GPUs.

A command example for starting a server using vLLM and accepting image and video inputs is as follows:

export MODEL=nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16

vllm serve $MODEL \
  --served-model-name nemotron-3.5-super-vl \
  --tensor-parallel-size 4 \
  --trust-remote-code \
  --gpu-memory-utilization=0.85 \
  --allowed-local-media-path=/ \
  --max-num-seqs=32 \
  --limit-mm-per-prompt '{"video": 1, "image": 1}' \
  --video-pruning-rate 0.5 \
  --media-io-kwargs '{"video": {"fps": 2, "num_frames": 256}}' \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_coder \
  --reasoning-parser nemotron_v3 \
  --speculative_config.method mtp \
  --speculative_config.num_speculative_tokens 3

As recommended parameters during inference, temperature 1.0 and top_p 0.95 are suggested for text generation, while temperature 0.6 and top_p 0.95 are suggested for visual tasks. Additionally, by controlling the enable_thinking argument of the chat template, you can toggle the reasoning (thinking) mode on and off.

Quantized and Converted Variants

→ Scroll horizontally to see all columns

Added Publisher Format Repository Smallest VRAM tier (build, est. memory)
2026-10-16 unsloth GGUF unsloth/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-GGUF Q3_K_XL 73.5GB (fits in 80GB VRAM)
2026-11-17 nvidia NVFP4 nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4 NVFP4 83.8GB (does not fit a single consumer GPU)
2026-11-17 nvidia NVFP4 nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4-LiLiCorr NVFP4 1.5GB (fits in 4GB VRAM)
2026-11-17 DevQuasar GGUF DevQuasar/nvidia.NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16-GGUF Q3_K_M 75.1GB (fits in 80GB VRAM)

File sizes of each build:

  • Available builds in unsloth/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-GGUF: IQ1_M 43.0GB / IQ2_XXS 47.6GB / IQ2_M 52.0GB / Q2_K_XL 53.8GB / IQ3_XXS 55.7GB / Q3_K_XL 61.2GB / Q4_K_XL 80.9GB / Q5_K_XL 103.1GB / Q6_K_XL 112.7GB / Q8_0 122.6GB
  • Available builds in nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4: NVFP4 69.9GB
  • Available builds in nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4-LiLiCorr: NVFP4 1.2GB
  • Available builds in DevQuasar/nvidia.NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16-GGUF: Q2_K 49.8GB / Q3_K_M 62.6GB / Q4_K_M 80.1GB / Q5_K_M 89.1GB / Q6_K 105.2GB / Q8_0 119.6GB

In addition, 32 converted build(s) from other uploaders exist on Hugging Face; this site lists only builds from the model’s publisher or established quantization maintainers.

This section is appended automatically by Local Model Watch when a converted build of this model appears after publication. Memory figures are estimated from the size of the distributed files. What the quantization names mean: glossary.

Related Articles

What to Read Next

Sources

Update History

  • 2026-11-17: Added converted builds to “Quantized and Converted Variants”: nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4, nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4-LiLiCorr, DevQuasar/nvidia.NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16-GGUF
  • 2026-11-17: Updated the hardware requirements table with the actual file sizes of unsloth/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-GGUF.
  • 2026-11-17: Changed the title to show what the article covers (VRAM requirements, file list, etc.).