Nvidia releases open Nemotron 3.5 Super VL model, now also on Amazon SageMaker JumpStart
The 120-billion-parameter vision-language model ships with open weights, datasets and training recipes, alongside two smaller Bolt models for local hardware. Nvidia's speed and accuracy figures remain vendor claims.
Key points
- 1Nemotron 3.5 Super VL is a 120B-parameter open vision-language model with a mixture-of-experts backbone and 12B active parameters, per AWS and Nvidia.
- 2The model takes images, video, documents and text as input, outputs text, and supports up to 1 million tokens of context with reasoning configurable per request.
- 3AWS said on Oct. 16 that the model can be deployed on Amazon SageMaker JumpStart in NVFP4 and BF16 variants via the console catalog or the Python SDK.
- 4Nvidia claims up to 5x faster throughput than the next best open model in its class and near-frontier accuracy at one-fifth the size; these figures are not independently verified in the sources.
- 5Two edge models accompany the release: Bolt 74B-A9B 24GB with 2-bit quantization is available now, and Bolt E4B 4GB is due Oct. 23.
Full story
Nvidia has released Nemotron 3.5 Super VL, an open vision-language model with 120 billion parameters that the company positions for always-on enterprise agents. According to Nvidia, the model ships with open weights, open datasets and training recipes, and is available on Hugging Face, through the ModelScope community and as an Nvidia NIM microservice. Two smaller edge models were announced at the same time: Nemotron 3.5 Bolt 74B-A9B 24GB, which is available now, and Nemotron 3.5 Bolt E4B 4GB, which Nvidia says will follow on Oct. 23.
A day after the launch, AWS said Nemotron 3.5 Super VL is available on Amazon SageMaker JumpStart. According to the AWS notice, the model uses a hybrid mixture-of-experts language backbone with 12 billion active parameters plus a vision encoder, accepts images, video and text as input, produces text output, and supports context lengths of up to 1 million tokens. AWS also describes a built-in multi-token prediction head for speculative decoding, reasoning that can be switched on or off per request, support for ten languages, structured JSON output and tool calling. SageMaker JumpStart offers two variants: an NVFP4 build meant for inference on G7 instances (RTX PRO 4500 Blackwell Server Edition GPUs), and BF16 reference weights, offered on ml.p5en.48xlarge, for fine-tuning, reinforcement learning, distillation and domain adaptation. Customers can deploy from the JumpStart model catalog in the SageMaker console or through the SageMaker Python SDK.
Nvidia's own performance statements have not been independently verified in the available sources. The company says Super VL reaches accuracy comparable to its largest reasoning model at roughly one-fifth the parameter count and claims it is 5x faster than the next best open agentic model in its class, citing throughput on the Artificial Analysis Intelligence Index leaderboard. Nvidia also says the model saw about 26 trillion training tokens and was then post-trained on roughly 3.8 million samples, followed by reinforcement learning in dozens of environments. The vision side targets optical character recognition, tables, charts and structured information in documents, web pages and PDFs, so that agents do not need to hand off between separate text and image systems.
Nvidia lists a number of organizations as customizing or integrating the model, including Atlassian for agentic search across Jira and Confluence, CrowdStrike for enterprise security blue teaming, Perplexity for search and agentic tasks, Salesforce for government workflow automation through Missionforce Operations, and Scale AI, which Nvidia says post-trained the model for finance and due diligence agents at Skydance and EY. The smaller Bolt 74B-A9B 24GB is designed, according to Nvidia, to run local agentic tasks on RTX and DGX Spark systems using a native 2-bit quantization that fits within 24GB of memory, with support for llama.cpp, Ollama, LM Studio, Unsloth and vLLM. Nvidia says Bolt E4B 4GB is aimed at conversational and physical agents on GeForce RTX, Jetson and RTX/DGX Spark hardware.
Why it matters
Open weights, training recipes and datasets give enterprises a path to fine-tune and host a multimodal agent model on their own infrastructure, and the SageMaker JumpStart listing lowers the barrier to trying it on AWS. The named integrators suggest early commercial uptake, but the speed and accuracy figures come only from Nvidia and should be read as vendor claims until independent testing is available.
Timeline
· Published
Topics#Nvidia#Nemotron#open models#vision-language models#AWS SageMaker#AI agents
Sources
This story draws on the following sources. Read them for full context.
