Nemotron 3.5 Super VL achieves similar accuracy to Nemotron 3 Ultra on agent tasks, according to NVIDIA
NVIDIA says Nemotron 3.5 Super VL achieves similar accuracy to Nemotron 3 Ultra in instruction following and tool use. It has approximately one-fifth as many parameters and is designed for enterprise agents and document analysis.
NVIDIA has published details on the performance of Nemotron 3.5 Super VL for enterprise agents. According to the company, it achieves similar accuracy to Nemotron 3 Ultra in instruction following and tool use, despite having approximately one-fifth as many parameters. This is a claim by the manufacturer, rather than independent confirmation of broadly comparable performance.
Nemotron 3.5 Super VL has 120 billion parameters in total and uses a Mixture-of-Experts architecture that activates 12 billion parameters for each token. It processes text, images, documents and video. It is designed for specialized agents running continuously in the background; when analyzing documents, it can connect information from a chart and a table without a separate vision model.
Alongside the weights, NVIDIA is publishing datasets and training procedures. The model is available on Hugging Face, and organizations can train it further on their own data and run it on their own infrastructure. The company also provides NeMo and Python procedures for customization.
The model is optimized for NVIDIA hardware. It uses the 4-bit NVFP4 format; on Blackwell-generation GPUs, computations also use 4-bit precision, while older cards store weights in 4 bits and compute in 16 bits. According to the company, the model achieves up to five times the throughput of other similarly sized open models on agent tasks using a single GPU.
Why it matters
For agents processing large volumes of documents, the combination of visual content analysis and high throughput may be important. However, the published comparisons demonstrate performance in specific tests; actual speed and accuracy on your own documents need to be verified.
Release card
Nemotron 3.5 Super VL
NVIDIA
- Specifications
- The model has 120 billion parameters in total; its Mixture-of-Experts architecture activates 12 billion parameters for each token.
- Inputs
- The model processes text, images, documents and video and generates text.
- Availability
- The model weights are publicly available on Hugging Face. The model can be trained further on your own data and run on your own infrastructure.
- CharXiv 80,8 % According to NVIDIA, the model scored 80.8 % on a benchmark focused on chart understanding.
- CharXiv s možností spouštět kód v jazyce Python 84 % According to NVIDIA, the model scored 84 % on a benchmark focused on chart understanding when able to run Python code.
- Continuously running enterprise agents that use tools and perform multi-step tasks.
- Document analysis that connects information from charts and tables.
- Adaptation using your own company data and deployment on your own infrastructure.
- On older GPUs, computations use 16-bit precision, even though the weights are stored in 4 bits.
According to NVIDIA, the model achieves similar accuracy to Nemotron 3 Ultra in instruction following and tool use, despite having approximately one-fifth as many parameters. The company distinguishes it from closed models by the ability to train it further on your own data and run it on your own infrastructure.
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
-
Super VL is designed for enterprise AI agents.; According to Nvidia, Super VL achieves similar accuracy to Nemotron 3 Ultra in instruction following and tool use.; Bolt models are designed to run locally on PCs with Nvidia RTX Spark and on DGX Spark workstations.
- Super VL is designed for enterprise AI agents.
- According to Nvidia, Super VL achieves similar accuracy to Nemotron 3 Ultra in instruction following and tool use.
- Bolt models are designed to run locally on PCs with Nvidia RTX Spark and on DGX Spark workstations.
-
According to the article, the models are available on Hugging Face.; The Bolt E4B 4GB variant is designed for conversational agents.; Nvidia says Super VL can complete tasks almost instantly.
- According to the article, the models are available on Hugging Face.
- The Bolt E4B 4GB variant is designed for conversational agents.
- Nvidia says Super VL can complete tasks almost instantly.
-
12 billion parameters are active per token.; The model processes text, images, documents and video.; Nvidia is publishing datasets and training procedures alongside the weights.; The model is optimized for Nvidia hardware.; A Bolt variant with 4 GB is planned for later for voice assistants.
- 12 billion parameters are active per token.
- The model processes text, images, documents and video.
- Nvidia is publishing datasets and training procedures alongside the weights.
- The model is optimized for Nvidia hardware.
- A Bolt variant with 4 GB is planned for later for voice assistants.
Two audiences, two different impacts
What this means
For individuals
A developer building a document agent can use a single model to connect information from tables, charts and text, which may simplify application development.
For a business
A company can adapt the model to its own data and run document processing on its own infrastructure. Optimization for NVIDIA hardware is an important factor in capacity planning.
ProcessesCheck the original
Event sources
confirmed by 3 independent sources · 4 publishers, 3 independent. We count feeds from the same owner only once.