Nemotron 3.5: What Nvidia’s Open Models Bring to the Workplace

Nvidia is bringing its open AI models closer to the workplace. On October 15, the company released Nemotron 3.5 Super VL for text and visual tasks alongside Bolt, a version designed for local use. The interesting part is less about another leaderboard position than the division of labor: companies can operate document agents themselves, while developers can use a compressed model on suitable hardware of their own.

Key takeaways

• Nemotron 3.5 Super VL processes text, images, and video and is designed to connect multistep tasks with tool calls.
• Bolt 74B-A9B 24GB is the compressed version for local agents. Its memory label does not replace checking the requirements of a specific runtime configuration.
• Model weights and technical instructions are available. The smaller Bolt E4B 4GB, however, is only scheduled for October 23.
• The performance claims come from Nvidia. Faster text generation does not automatically mean a finished assignment will be completed correspondingly faster.

One model for documents and the work that follows

Super VL combines language processing with visual understanding. This allows it to work with tables, charts, and page layouts rather than relying solely on text that has already been extracted. That matters for office tasks: a scanned receipt presents information differently from a neatly labeled database. An agent needs to recognize the content before it can compare amounts or prepare a record.

The model card lists 120 billion total parameters, with 12 billion active per token. Parameters are the numerical values learned during training; a token is a unit of text. In a mixture-of-experts architecture, only some of the available experts are activated at a time. That reduces computation compared with using every parameter, but it does not turn the large model into an ordinary small desktop application.

The distinction between understanding and execution matters. In Nvidia’s tool-calling example, the model requests a function; the surrounding application executes it and returns the result. Access to inventory or business systems therefore comes from the software connected to the model. Downloading a model does not automatically give it access to a company’s operations.

For businesses, that connection is where the value could emerge: an application could read a supplier invoice, look up an internal purchase order number, and flag discrepancies for review. This is a potential use case, rather than a workflow we have tested. The expansion of Amp beyond programming also illustrates how agents are moving from individual chats toward complete assignments.

What Bolt enables locally, and what 24 GB does not promise

Bolt 74B-A9B 24GB is produced by compressing the larger model and has 74 billion total parameters, with nine billion active. Nvidia combines this reduction with substantially lower numerical precision for large portions of the model weights. In simple terms, the learned values are stored more compactly; additional training is intended to recover capabilities that compression can reduce.

The GGUF version is an interesting starting point. GGUF is a format used by local model runtimes. The official model page describes running it with llama.cpp and also lists Ollama. Nvidia thus provides a documented download and integration path. That does not guarantee that every installed version of these applications already supports the new model type equally well.

The memory figure in particular deserves a closer look. The GGUF documentation includes configurations for 28 and 32 GB and lists 32 GB or more for discrete RTX graphics cards. Depending on the recipe, a helper model, visual processing, and cached conversation data add to the requirements. The “24GB” name therefore does not establish that any graphics card with exactly that capacity can run every documented configuration.

Anyone interested in trying Bolt should first match the intended runtime with its memory requirements. A manageable task with known results is a useful next step, such as extracting information from nonsensitive sample documents. Connecting additional tools becomes worthwhile once that works reliably. Running the model locally can replace cloud inference; a tool that searches the web still needs a network connection.

Open weights create choices, rather than automatic savings

The models use the OpenMDW-1.1 license. It permits broad use and modification; redistribution requires retaining the license and applicable notices of origin, among other conditions. It also includes a warranty disclaimer. In practical terms, available weights make customization and self-hosting easier, but they do not perform the integration or assume responsibility for a particular application.

Performance figures also need context. Nvidia advertises up to five times the throughput for Super VL compared with other open models in its size class. The technical article discusses token output among its measurements. A complete agent assignment also involves fetching documents, calling tools, and potentially recovering from failed attempts. The comparison therefore does not establish five times the overall working speed or equivalent savings for every project.

Self-hosting changes the cost structure as well. Hardware, maintenance, and operational support can replace a straightforward usage fee. Whether this is cheaper depends on utilization and the task. For now, the stronger reason to consider the new options is the ability to adapt a model to an organization’s workflows and arrange its operation directly.

A useful starting point is a specific task

Super VL is aimed primarily at teams that want to connect visual information with business tools. Bolt offers a smaller local starting point for technically experienced users with suitable hardware. The E4B model, which has only been announced, should not yet be treated as an available product when making that decision.

The release shifts a practical boundary: an open agent model does not have to remain exclusively in a large data center. Whether it becomes a useful assistant at someone’s desk depends on a clearly defined task, sufficient memory headroom, and results that can be checked. Those are more revealing places to look for value than the next general leaderboard.