NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4: ~84GB Memory

At a Glance
| Item | Value |
|---|---|
| Repository | nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4 |
| 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 has released an NVFP4 quantized version of its vision-language model, Nemotron 3.5 Super VL, named NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4. This open model supports reasoning, coding, tool use, document comprehension, and video understanding, and reduces memory requirements through Quantization-Aware Distillation from the original BF16 version. It features a hybrid architecture combining Mamba-2, Mixture-of-Experts (MoE), and attention layers, and also supports speculative decoding via multi-token prediction (MTP).
Specifications
- Total parameters: 120B
- Active parameters: 12B
- Architecture: Mamba-2–Transformer hybrid Mixture of Experts (512 routed experts per MoE layer, C-RADIO vision encoder, vision-language connector, equipped with Multi-Token Prediction)
- Context length: Up to 1M tokens
- Input modalities: Text, images (RGB, JPEG/PNG), video (MP4, up to 2 minutes)
- Output modality: Text
Performance
Benchmark comparison results between the original model (BF16 version) and this quantized version (NVFP4 version) as stated in the model card are shown below. These figures were compared under unified conditions such as evaluation setups and harnesses.
| Benchmark | Nemotron 3.5 Super VL BF16 | Nemotron 3.5 Super VL NVFP4 |
|---|---|---|
| GPQA-Diamond | 89.52 | 89.90 |
| HLE, text-only | 31.79 | 31.44 |
| MMLU-ProX | 76.91 | 73.76 |
| AA-Omniscience accuracy | 26.20 | 25.86 |
| AA-Omniscience hallucination rate | 30.0 | 27.7 |
| AA-Omniscience index | 4.03 | 5.3 |
| APEX shortlist | 88.36 | 89.36 |
| LiveCodeBench v6 | 88.22 | 87.41 |
| SciCode, subtask | 47.13 | 46.61 |
| AA-LCR v1.1 | 68.62 | 67.69 |
| GraphWalks | 89.71 | 90.34 |
| IFBench | 70.86 | 71.50 |
| LMArena v3 proxy | 52.67 | 51.93 |
| WMT24++ | 84.80 | 84.76 |
| tau2-bench retail | 69.24 | 69.30 |
| tau2-bench airline | 76.50 | 75.75 |
| tau2-bench telecom | 73.19 | 73.19 |
| tau3-bench banking | 18.88 | 17.78 |
| GDPval-AA v2 | 1211 | 1206 |
| BrowseComp | 78 | 77.5 |
| SWE-bench Verified | 71.10 | 71.07 |
| SWE-bench Multilingual | 67.72 | 67.00 |
| SWE-bench Pro | 51.78 | 51.30 |
| Terminal-Bench 2.1 | 66.70 | 60.80 |
| BabyVision | 29.60 | 29.51 |
| MMLongBench-Doc | 62.82 | 62.82 |
| V* | 85.60 | 85.95 |
| OCR-Reasoning | 65.08 | 65.09 |
| MathVision | 78.16 | 77.79 |
| HLE, full set | 29.20 | 29.00 |
| OCRBench v2, EN | 68.12 | 68.34 |
| OCRBench v2, ZH | 59.15 | 59.03 |
| CharXiv reasoning | 80.82 | 79.88 |
| Video-MME v2 | 32.85 | 32.54 |
| OSWorld-Verified | 51.14 | 52.3 |
| BrowseComp [offline] | 78 | 77.5 |
| WebVoyager [offline] | 77.72 | 66.31 |
| AA-Briefcase-Lite [offline] | 1137 (-) | 1252 (-) | 40 | 1137 (-) | 1188 (-) | 45 |
| AutomationBench-AA [offline] | 32.44 | 32.3 |
| GDP.pdf [offline] | 6.6 (AA 6.6) | 6.0 |
| Terminal Bench 4 [offline] | 1 | 1.5 |
| DeepSWE [offline] | 24.65 | 18.2 |
These scores are measured and published by NVIDIA, the model creator. As can be seen from the table, in some tasks such as GPQA-Diamond handling PhD-level difficult questions (89.52 vs. 89.90) and OSWorld-Verified measuring the ability to operate actual OS screens (51.14 vs. 52.3), the quantized version records slightly higher scores. Furthermore, it maintains high accuracy comparable to the BF16 version in SWE-bench Verified (71.10 vs. 71.07) measuring problem-solving capabilities on real repositories, LiveCodeBench v6 (88.22 vs. 87.41) using new competitive programming problems, and OCRBench v2 EN (68.12 vs. 68.34) measuring text recognition in images.
On the other hand, scores drop from 66.70 to 60.80 on Terminal-Bench 2.1, which evaluates the ability to complete tasks by executing commands on a terminal, and non-negligible drops in accuracy are observed in some agent-based and specific tasks such as WebVoyager [offline] (from 77.72 to 66.31) and DeepSWE [offline] (from 24.65 to 18.2). Overall, while advanced reasoning, code generation, and visual understanding capabilities are preserved, it indicates that trade-offs due to quantization exist in complex autonomous environment operation tasks.
Strengths and Use Cases
This model is designed primarily for building multimodal agents responsible for coding, search, tool-assisted workflows, document question answering, and image/video understanding.
Specifically, it can accept not only still images but also MP4 videos of up to 2 minutes directly as input, processing visual information via the C-RADIO vision encoder. It can process multi-page documents as a combination of page images and text, and leverages a massive context length of up to 1M tokens to easily continue working while retaining lengthy source code bases, past dialogue history, and tool execution results all at once.
It also features a function to toggle the reasoning mode on and off via chat template arguments (enable_thinking=True/False). This allows flexible operation, such as enabling thinking for difficult tasks requiring deep reasoning steps and disabling it for dialogues requiring fast responses. Supported language tags include English, Japanese (ja), Spanish, French, German, and Italian.
Hardware Requirements
Estimated requirements (calculated by Local Model Watch)
| Your VRAM | Quantization | File size | Est. memory needed |
|---|---|---|---|
| More than 84GB of VRAM (multi-GPU or CPU offload required) | NVFP4 | 69.9GB | 83.8GB |
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. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.
Can You Run It Locally?
Not usable in Ollama, LM Studio and llama.cpp yet — we have found no GGUF build.
The publisher ships safetensors only. However, llama.cpp’s registry does list this architecture, so conversion to GGUF is possible and the model will run once someone publishes a converted build. 12 converted build(s) from other uploaders exist. Today it can be run with transformers or vLLM, using the memory figures in the table above.
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).
How to Get It
The model weights are provided in safetensors format, and the OpenMDW-1.1 license applies. Both commercial and non-commercial uses are permitted, and they can be downloaded directly without going through application or consent procedures (Gated approval) on Hugging Face.
As for inference engines, vLLM, SGLang, and TensorRT-LLM are officially supported, and guidance for llama.cpp and Unsloth is also provided. The target hardware is the NVIDIA Hopper or Blackwell generation, with inference executed as W4A4 in Blackwell environments and W4A16 on other platforms.
Models can be downloaded via the Hugging Face CLI.
hf download nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4 --local-dir./models/super-vl-nvfp4
When launching a 1M context-capable server using vLLM on two H100 GPUs (tensor parallel size 2), run the following command on the official Docker container image (vllm/vllm-openai:nightly-3f1bcfd82bb2e93fcd364fc7a5491718453a77e2 or vllm/vllm-openai:v0.32.0).
export MODEL=nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4
vllm serve $MODEL \
--served-model-name nemotron-3.5-super-vl \
--tensor-parallel-size 2 \
--max-num-seqs 128 \
--trust-remote-code \
--gpu-memory-utilization 0.89 \
--max-num-batched-tokens 16384 \
--block-size 64 \
--prefix-match-unit 64 \
--compilation_config.max_cudagraph_capture_size 256 \
--mamba-backend flashinfer \
--mamba-ssm-cache-dtype float16 \
--enable-mamba-cache-stochastic-rounding \
--mamba-cache-philox-rounds 5 \
--enable-mamba-shared-prefix-checkpoint \
--mamba-cache-mode align \
--speculative-config.method mtp \
--speculative-config.num_speculative_tokens 3 \
--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
Additionally, for DGX Spark environments, a serving configuration combining a dedicated speculative drafter, LiLiCorr, is also provided to accelerate generation.
Quantized and Converted Variants
→ Scroll horizontally to see all columns
| Added | Publisher | Format | Repository | Smallest VRAM tier (build, est. memory) |
|---|---|---|---|---|
| 2026-11-17 | nvidia | NVFP4 | nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4-LiLiCorr | NVFP4 1.5GB (fits in 4GB VRAM) |
File sizes of each build:
- Available builds in nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4-LiLiCorr: NVFP4 1.2GB
In addition, 12 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
- NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16: 80GB+ VRAM, GGUF Builds
- DeepSeek-V4-Pro-0813-nvfp4-DSpark: ~1005GB Memory
- NVIDIA Releases Dory1.0-4B: A Recurrent 4B LLM with Dynamic Loops
- Muse-Spark-Open-816B-A42B-DFlash2 Vision-Language Model: 12GB+ VRAM
What to Read Next
- Find models by VRAM (This model does not fit even in 80GB; its smallest build needs about 84GB) → VRAM quick reference, including models over 80GB
- Explore the same model family → NVIDIA-Nemotron-3.5-Super-VL family overview (2 articles, 4 converted builds)
- Engines that run this model → vLLM
- What NVFP4 mean and where to get this model → NVFP4 / MXFP4 format guide and models / Quantization and model-format glossary
- Learn about the publisher → NVIDIA: models, licenses and articles
Sources
- nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4 (Hugging Face)
- NVIDIA Nemotron 3.5 Super VL and Lightning Technical Report
- OpenMDW-1.1 License
Update History
- 2026-11-17: Added converted builds to “Quantized and Converted Variants”: nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4-LiLiCorr

