Nvidia's Nemotron 3.5 Super VL model is now available in Amazon SageMaker JumpStart
Nvidia's Nemotron 3.5 Super VL model is now available in Amazon SageMaker JumpStart, in NVFP4 (75 GB) and BF16 (249 GB) variants with a context of up to 1 million tokens. The model, with 120 billion parameters, was previously released as open for enterprise agents and document analysis.
Nvidia's Nemotron 3.5 Super VL model is now available on the Amazon SageMaker JumpStart platform, where it can be deployed directly from SageMaker Studio or via the Python SDK. Two variants are available: the quantized NVFP4 at approximately 75 GB, and the full BF16 at approximately 249 GB. The model supports a context window of up to 1 million tokens (JumpStart configurations use a range of 520 to 15 600 tokens) and includes a built-in Multi-Token Prediction head for speculative decoding. It is intended for document analysis, GUI understanding for computer-use agents, answering questions about video, and other multimodal agentic tasks. It is licensed under the OpenMDW License Agreement version 1.1.
Nvidia first introduced Nemotron 3.5 Super VL as an open multimodal model with 120 billion parameters, of which 12 billion are active per token (Mixture-of-Experts architecture). The model combines a language core with the C-RADIO vision encoder and processes text, images, documents, and video. According to Nvidia, Super VL achieves accuracy similar to the larger Nemotron 3 Ultra model in instruction-following and tool-use tasks, while having roughly one-fifth the number of parameters.
At the same time, Nvidia also released two smaller Bolt-series models for local deployment: Nemotron 3.5 Bolt 74B-A9B, 24 GB in size, for agentic tasks on workstations with Nvidia RTX Spark and DGX Spark chips, and the smaller Nemotron 3.5 Bolt E4B, 4 GB in size, intended for conversational agents. For all models, Nvidia releases not only the weights but also the datasets and training procedures; the models are available on the Hugging Face platform.
Why it matters
Availability in Amazon SageMaker JumpStart lowers the deployment barrier for companies that want to run the model on AWS without building their own GPU infrastructure, with clearly defined instances (ml.g7 for NVFP4, ml.p5en for BF16) and two tiers balancing latency and accuracy. For developers and companies, this represents an open alternative to closed models for agentic tasks and document work, one that can be fine-tuned on their own data and operated outside the terms of a closed-model provider.
Release card
Nemotron 3.5 Super VL
Nvidia
- Specifications
- 120 billion parameters total, 12 billion active per token (hybrid Mixture-of-Experts architecture).
- Context
- Up to 1 million tokens; the configuration in Amazon SageMaker JumpStart uses a range of approximately 520 to 15 600 tokens.
- Inputs
- Input: text, images, and video; output: text.
- Licence
- OpenMDW License Agreement, verze 1.1.
- Availability
- The model is available as open weights on the Hugging Face platform and can also be deployed via the Amazon SageMaker JumpStart service (SageMaker Studio or Python SDK).
- CharXiv reasoning (varianta BF16) 80,82 % The figure indicates the model's measured success rate on the task of interpreting charts and images with reasoning enabled.
- OSWorld-Verified (varianta NVFP4) 52,30 % The figure indicates the model's measured success rate on the task of an agent visually controlling a computer's graphical interface.
- Výstupní propustnost na instanci ml.p5en.48xlarge (BF16) 140,85 tokenů za sekundu na požadavek The figure indicates the measured average text generation speed with five concurrent requests in the given configuration.
- document analysis and understanding of charts, tables, and multi-page PDFs
- agents controlling a computer's graphical interface (computer-use agents)
- answering questions about video as part of multimodal agentic tasks
- the configuration in Amazon SageMaker JumpStart does not support deploying LoRA adapters
- full BF16 precision requires an instance with 1 128 GB of GPU memory (8x Nvidia H200)
Nvidia compares the model to its own larger model Nemotron 3 Ultra, whose accuracy it is said to match with roughly one-fifth the number of parameters, and claims that when fine-tuned for specific agentic tasks, it completes tasks faster than closed models from competing companies.
The card summarizes information from the article and any dated corrections, with a link to the original source. It is not our assessment of the model. It does not yet have a dedicated editorial profile. Model selection and other announcements →
What was added since the original report
Verified updates
-
The model is now available in Amazon SageMaker JumpStart; Supports context length of up to 1 million tokens; There are two variants: NVFP4 (75 GB, quantized) and BF16 (249 GB, full precision); The model includes a Multi-Token Prediction head for speculative decoding; Specific use cases include document intelligence, GUI understanding for computer-use agents, and video question answering
- The model is now available in Amazon SageMaker JumpStart
- Supports context length of up to 1 million tokens
- There are two variants: NVFP4 (75 GB, quantized) and BF16 (249 GB, full precision)
- The model includes a Multi-Token Prediction head for speculative decoding
- Specific use cases include document intelligence, GUI understanding for computer-use agents, and video question answering
-
Super VL is intended for enterprise AI agents.; According to Nvidia, Super VL achieves accuracy similar to Nemotron 3 Ultra in instruction-following and tool use.; The Bolt models are intended for local deployment on PCs with Nvidia RTX Spark and DGX Spark workstations.
- Super VL is intended for enterprise AI agents.
- According to Nvidia, Super VL achieves accuracy similar to Nemotron 3 Ultra in instruction-following and tool use.
- The Bolt models are intended for local deployment on PCs with Nvidia RTX Spark and DGX Spark workstations.
-
According to the article, the models are available on the Hugging Face platform.; The Bolt E4B 4GB variant is intended for conversational agents.; For Super VL, Nvidia states the ability to complete tasks almost instantly.
- According to the article, the models are available on the Hugging Face platform.
- The Bolt E4B 4GB variant is intended for conversational agents.
- For Super VL, Nvidia states the ability to complete tasks almost instantly.
-
12 billion parameters are active per token.; The model processes text, images, documents, and video.; Alongside the weights, Nvidia also releases the datasets and training procedures.; The model is optimized for Nvidia hardware.; The 4 GB Bolt variant is planned for later, for voice assistants.
- 12 billion parameters are active per token.
- The model processes text, images, documents, and video.
- Alongside the weights, Nvidia also releases the datasets and training procedures.
- The model is optimized for Nvidia hardware.
- The 4 GB Bolt variant is planned for later, for voice assistants.
Two audiences, two different impacts
What this means
For individuals
Developers can try deploying the multimodal agentic model via Amazon SageMaker JumpStart or Hugging Face and choose between the faster quantized NVFP4 variant and the full BF16 variant depending on available memory and required precision.
For a business
Companies can deploy an open multimodal model for agents and document analysis on AWS infrastructure without depending on a closed provider, with the option to fine-tune it on their own data; the choice between the NVFP4 and BF16 variant directly affects cloud compute costs and GPU instance requirements.
DevelopmentCheck the original
Event sources
confirmed by 3 independent sources · 5 publishers, 3 independent. We count feeds from the same owner only once.