Nemotron 3.5 Super VL processes documents and video and activates 12 billion parameters per token
Nemotron 3.5 Super VL has 120 billion parameters, of which 12 billion are active per token, according to NVIDIA. It processes text, images, documents and video. Alongside the weights, the company is also releasing datasets and training procedures.
Details of Nemotron 3.5 Super VL describe a Mixture-of-Experts architecture that, according to NVIDIA, activates 12 out of 120 billion parameters per token. The model processes text, images, documents and video. Alongside the weights, the company is also releasing datasets and training procedures; the model is optimized for NVIDIA hardware.
NVIDIA has released this open model for enterprise agents and document analysis. According to the company, it can interpret a chart, connect it to a table elsewhere in the document and summarize both without a separate vision model. It offers businesses deployment on their own servers and customization with their own data. It provides the NeMo framework and Python tools for customization.
The release is accompanied by Nemotron 3.5 Bolt 74B-A9B 24GB, a model for local use that, according to the article, requires at least 24 GB of memory and uses 2-bit weight quantization. It supports the tools llama.cpp, Ollama, LM Studio and vLLM. The press release also introduces Nemotron 3.5 Bolt E4B 4GB; the article from The Decoder clarifies that this variant for voice assistants is not due until later in the month.
Why it matters
Analyzing charts, tables and multi-page documents with a single model could simplify building an agent to work with company materials. However, the manufacturer is making its own claims about the model's capabilities; their usefulness for specific documents needs to be verified.
Release card
Nemotron 3.5 Super VL
NVIDIA
- Specifications
- The model has 120 billion parameters, of which 12 billion are active per token.
- Inputs
- The model processes text, images, documents and video and generates text.
- Availability
- NVIDIA has released the model weights for deployment on your own servers and customization with your own data. It also provides datasets and training procedures.
- CharXiv 80,8 % According to NVIDIA, the model scored 80.8 % in a chart comprehension evaluation.
- CharXiv s možností spouštět kód v jazyce Python 84 % According to NVIDIA, the score in a chart comprehension evaluation reaches 84 % if the model can also execute Python code.
- Analysis of tables, charts and multi-page documents.
- Continuous operation of enterprise agents that use tools and handle multi-step tasks.
- Customization with your own data and deployment on your own servers.
NVIDIA presents the model as an alternative to leading closed models for businesses that want to run the model on their own servers and customize it with their own data.
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
-
12 billion parameters are active per token.; The model processes text, images, documents and video.; Alongside the weights, Nvidia is also releasing datasets and training procedures.; The model is optimized for Nvidia hardware.; The Bolt variant with 4 GB is planned for later release for voice assistants.
- 12 billion parameters are active per token.
- The model processes text, images, documents and video.
- Alongside the weights, Nvidia is also releasing datasets and training procedures.
- The model is optimized for Nvidia hardware.
- The Bolt variant with 4 GB is planned for later release for voice assistants.
Two audiences, two different impacts
What this means
For individuals
For an individual who wants to run an agent on their own computer, the requirement for at least 24 GB of memory is a key condition when choosing an available variant.
For a business
The ability to customize the model with company data and run it on company servers opens up another infrastructure option for continuous document processing. Optimization for NVIDIA hardware is an important consideration when assessing this option.
ProcessesCheck the original
Event sources
independently confirmed · 2 publishers, 1 independent. We count feeds from the same owner only once.