Nvidia’s 120B vision model runs on 12B active parameters, now one click away on AWS

Amazon Web Services has added Nvidia’s Nemotron 3.5 Super VL to SageMaker JumpStart. The vision-language model has 120B parameters, of which only 12B are active per token, and it supports context windows of up to 1 million tokens.

Multimodal input and agent features in one model

Nemotron 3.5 Super VL takes images, video and text as input and produces text. According to AWS, it is aimed at document intelligence, GUI understanding, video question answering and multimodal agent applications. Its language backbone is a hybrid Mixture-of-Experts design, paired with a vision encoder.

Several features are built for agent work. A Multi-Token Prediction head supports speculative decoding to speed up text generation, and reasoning can be switched on or tuned for each request. The model also returns structured JSON and supports tool calling. It covers ten languages, among them English, French, Spanish, German, Japanese, Hindi, Korean and Chinese.

On training, AWS describes a two-stage recipe. The model was first fine-tuned under supervision on mixed visual material, including photos, clips, documents, charts and screenshots of interfaces, and was then refined with GRPO, a reinforcement-learning method.

JumpStart offers two variants:

• NVFP4, tuned for inference on G7 instances with NVIDIA RTX PRO 4500 Blackwell Server Edition GPUs.
• BF16 reference weights, meant for customization (SFT, RL, distillation and domain adaptation) on ml.p5en.48xlarge instances.

Customers can pick the model in the JumpStart catalog in the SageMaker console or deploy it with the SageMaker Python SDK. The announcement does not give pricing or a list of supported AWS Regions.

Where the model comes from

The AWS post is a distribution step. Nvidia released Nemotron 3.5 Super VL as an open model on October 15, 2026, according to Interesting Engineering, alongside a smaller model called Bolt 74B-A9B. That report says Super VL is also available through Hugging Face, OpenRouter, ModelScope and Nvidia NIM. Nvidia’s own wording, as quoted there, describes an open model that “combines reasoning and tool use with text, image, document, and video understanding”. Nvidia says it finishes tasks about five times faster than the nearest competitor in its class, a company claim that has not been independently verified.

Adopters named in that report include Scale AI, which adapted the model for finance and due-diligence agents, and Mercor, which post-trained it for coding tasks. The same report says the license terms were not spelled out.

Hybrid design behind the 12B active parameter efficiency

The interesting part is the efficiency trade, and the new facts around it show why. Nvidia’s Megatron Bridge documentation describes the model as a Nemotron-H hybrid decoder combined with a RADIO vision encoder. Hybrid designs of this kind aim to keep long-context cost down, which matters when a 1M-token window has to hold hours of video frames or hundreds of document pages. Interesting Engineering reports that the model was pretrained on about 26 trillion tokens and post-trained on roughly 3.8 million samples, with reinforcement learning across dozens of environments.

Activating 12B of 120B parameters is what makes the speed numbers plausible. Artificial Analysis, as reported by Let’s Data Science, scored the model 26.5 on its Intelligence Index and measured roughly 630 output tokens per second with 100,000-token prompts. That test used a pre-release endpoint, so the figure may shift on production hardware such as the G7 instances AWS names.

The release also shows how AWS is positioning JumpStart. A day earlier, AWS published a guide to Meta’s Muse Spark Open on the same service. That model has 816B total parameters, 42B active, and runs on far larger ml.p5e and ml.p6-b300 instances. Nemotron 3.5 Super VL sits at the other end, with an NVFP4 build meant for a single-GPU-class G7 target. For teams that want a multimodal agent model in their own AWS account, that gap in hardware requirements is a practical difference.

Nvidia has been releasing open weights at a steady pace. Its Dory 4B model, covered earlier this week, is another example. Letting customers run such models inside their own cloud accounts speaks to the data-control argument Nvidia makes for open models. Early testing by independent parties will still matter: Let’s Data Science cautions that benchmarks are no substitute for application-specific checks on accuracy, tool use, latency and cost.