Can Nemotron 3.5 Super VL run on a single machine at home?

Nemotron 3.5 Super VL needs 4× H100 or 2× GB200 GPUs for its BF16 checkpoint, according to the Hugging Face model card. For local use NVIDIA released Nemotron 3.5 Bolt 74B-A9B on the same day: a 74B model with 9B active parameters, quantized to about 2 bits so the whole package fits in 24GB on cards such as the RTX 5090 or DGX Spark.

How does Nemotron 3.5 Super VL compare to Qwen and GLM models?

NVIDIA's own table puts Nemotron 3.5 Super VL at 89.52 on GPQA Diamond and 71.10 on SWE-bench Verified, close to Qwen3.5-122B-A10B. It trails Qwen3.8-Flash-Next and GLM-5.3-Flash on coding and agent tests: on Terminal-Bench 2.1 Super VL scores 60.67 against 85.80 and 82.70. NVIDIA's pitch is speed, with up to 5x the single-GPU agent throughput of open models its size.

Is Nemotron 3.5 Super VL free for commercial use?

Nemotron 3.5 Super VL is published under the OpenMDW-1.1 license, the same permissive license NVIDIA uses for the rest of the Nemotron 3.5 family. The weights are on Hugging Face and ModelScope in BF16 and NVFP4 versions, and NVIDIA also released training datasets, including 40 vision-language datasets, plus NeMo Gym reinforcement-learning environments.

Where can I try Nemotron 3.5 Super VL without downloading it?

NVIDIA lists hosted access to Nemotron 3.5 Super VL through OpenRouter and as an NVIDIA NIM microservice, in addition to open weights on Hugging Face and ModelScope. For self-hosting, the model card gives vLLM and SGLang commands, including a multi-token-prediction speculative decoding setup that serves up to 1M tokens of context on four H100 GPUs.

What is the smaller Nemotron 3.5 Bolt E4B model?

Nemotron 3.5 Bolt E4B is a third, 4GB member of the family that NVIDIA announced alongside Super VL and Bolt 74B-A9B. NVIDIA aims it at fast voice assistants, interactive characters and physical agents on GeForce RTX, Jetson and DGX Spark hardware, and says it arrives on October 23 rather than at launch.

NVIDIA · 2026-10-15 · major

Nemotron 3.5 Super VL — NVIDIA's open 120B vision model for agents

Nemotron 3.5 Super VL is NVIDIA's open 120B vision-language MoE with 12B active parameters and a 1M-token context. It shipped on 2026-10-15 with Nemotron 3.5 Bolt 74B-A9B, a 2-bit model that fits in 24GB.

NVIDIA Nemotron 3.5 Super VL agentic AI announcement graphic

NVIDIA's new open model reads text, documents, charts and short videos in one checkpoint, with only 12B parameters active per token.

Key specs

SWE-bench	71.10%
GPQA	89.52%

Quick facts

Maker	NVIDIA
Parameters	120B total, 12B active
Context window	Up to 1M tokens
Inputs	Text, images (JPEG, PNG), MP4 video up to 2 minutes
License	OpenMDW-1.1
Hardware (BF16)	4× H100 or 2× GB200
Sibling model	Nemotron 3.5 Bolt 74B-A9B, 24GB, 256K context

Benchmarks

SWE-bench Verified

Nemotron 3.5 Super VL		71.1%
Qwen3.5-122B-A10B		71.9%
Nemotron 3 Ultra		73.9%
DeepSeek-V4-Flash0731		75.4%
Qwen3.8-Flash-Next		78.7%
GLM-5.3-Flash		79.1%

source ↗

Terminal-Bench 2.1

Nemotron 3.5 Super VL		60.67%
Qwen3.5-122B-A10B		48.7%
Nemotron 3 Ultra		54.8%
DeepSeek-V4-Flash0731		76.8%
Qwen3.8-Flash-Next		85.8%
GLM-5.3-Flash		82.7%

source ↗

What is it?

Nemotron 3.5 Super VL adds vision to NVIDIA's open Nemotron 3.5 line: one 120B checkpoint handles reasoning, tool use, OCR, charts, PDFs and MP4 video up to two minutes. NVIDIA released it on 2026-10-15 together with Nemotron 3.5 Bolt 74B-A9B, a compressed version for 24GB laptops and desktop GPUs.

How does it work?

The architecture mixes Mamba-2 state-space layers, attention and a mixture-of-experts router, so only 12B of the 120B parameters run per token; a C-RADIO vision encoder feeds images and video in. NVIDIA distilled it from Nemotron 3 Ultra and trained it on about 26 trillion tokens before post-training and reinforcement learning. Bolt goes further with 2-bit experts recovered through quantization-aware distillation.

Why does it matter?

Document-heavy agents usually need a separate vision model next to the language model. With an open license, published training data and up to 5x the single-GPU agent throughput NVIDIA claims over open rivals, teams can self-host one model for invoices, receipts and screen recordings. NVIDIA names Scale AI and Ramp among its post-training partners, and Atlassian, CrowdStrike, Perplexity and Salesforce among the platforms building it in.

Who is it for?

agent builders, document-AI teams, local-LLM users with 24GB GPUs

Frequently asked questions

Can Nemotron 3.5 Super VL run on a single machine at home?: Nemotron 3.5 Super VL needs 4× H100 or 2× GB200 GPUs for its BF16 checkpoint, according to the Hugging Face model card. For local use NVIDIA released Nemotron 3.5 Bolt 74B-A9B on the same day: a 74B model with 9B active parameters, quantized to about 2 bits so the whole package fits in 24GB on cards such as the RTX 5090 or DGX Spark.
How does Nemotron 3.5 Super VL compare to Qwen and GLM models?: NVIDIA's own table puts Nemotron 3.5 Super VL at 89.52 on GPQA Diamond and 71.10 on SWE-bench Verified, close to Qwen3.5-122B-A10B. It trails Qwen3.8-Flash-Next and GLM-5.3-Flash on coding and agent tests: on Terminal-Bench 2.1 Super VL scores 60.67 against 85.80 and 82.70. NVIDIA's pitch is speed, with up to 5x the single-GPU agent throughput of open models its size.
Is Nemotron 3.5 Super VL free for commercial use?: Nemotron 3.5 Super VL is published under the OpenMDW-1.1 license, the same permissive license NVIDIA uses for the rest of the Nemotron 3.5 family. The weights are on Hugging Face and ModelScope in BF16 and NVFP4 versions, and NVIDIA also released training datasets, including 40 vision-language datasets, plus NeMo Gym reinforcement-learning environments.
Where can I try Nemotron 3.5 Super VL without downloading it?: NVIDIA lists hosted access to Nemotron 3.5 Super VL through OpenRouter and as an NVIDIA NIM microservice, in addition to open weights on Hugging Face and ModelScope. For self-hosting, the model card gives vLLM and SGLang commands, including a multi-token-prediction speculative decoding setup that serves up to 1M tokens of context on four H100 GPUs.
What is the smaller Nemotron 3.5 Bolt E4B model?: Nemotron 3.5 Bolt E4B is a third, 4GB member of the family that NVIDIA announced alongside Super VL and Bolt 74B-A9B. NVIDIA aims it at fast voice assistants, interactive characters and physical agents on GeForce RTX, Jetson and DGX Spark hardware, and says it arrives on October 23 rather than at launch.

Try it

Pull `nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4` from Hugging Face, or `docker model run hf.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB`