In brief
- Nemotron 3.5 Super VL is a 120-billion-parameter multimodal model available with open weights.
- NVIDIA says Bolt 74B-A9B 24GB was designed to fit in 24 GB of memory for local agent workloads.
- The 4GB Bolt E4B model is due on 23 October and is designed for GeForce RTX, Jetson and RTX/DGX Spark.
- Super VL can run on DGX Spark, DGX Station, RTX systems, Jetson, in data centers and in the cloud.
On 15 October 2026, NVIDIA expanded the Nemotron 3.5 family with a new open model and two options for edge devices. The company announced the 120-billion-parameter Nemotron 3.5 Super VL and Nemotron 3.5 Bolt 74B-A9B, both of which are now available; the Bolt E4B model is scheduled to launch on 23 October. NVIDIA is positioning the models for a range of agent deployments, from enterprise servers to local hardware and edge devices.
What tasks does Super VL focus on?
According to NVIDIA, Super VL was developed for multimodal agent tasks such as instruction following, tool use, multi-step workflows and visual content understanding. The model is designed to process text, tables and charts in web pages, PDFs and visual documents. This allows an agent to read document content and use the information it extracts in subsequent steps.
The company says Super VL is the fastest among open agent models in a comparable class and delivers up to 5 times the throughput of other models. These performance claims are based on NVIDIA’s own statements. The model’s open weights, datasets and training recipes are being released, allowing organizations to fine-tune it on their own data and run it on their own infrastructure.
NVIDIA says Super VL was trained on approximately 26 trillion tokens, with the final training stage using approximately 3.8 million examples and reinforcement learning. NVIDIA NeMo and Python-based workflows are available for customization. The model is accessible through Hugging Face, Open Router, ModelScope and NVIDIA NIM.
Bolt models focus on local execution
Nemotron 3.5 Bolt 74B-A9B 24GB was designed to fit in 24 GB of memory using the native 2-bit quantization method specified by NVIDIA. According to the company, the model can handle agent tasks such as web research, writing and code debugging on local systems without sending them to the cloud. Available through Hugging Face, the model supports llama.cpp, Ollama, LM Studio, Unsloth and vLLM.
The smaller Bolt E4B 4GB is positioned for use cases such as fast speech agents and local voice assistants. NVIDIA is targeting GeForce RTX, Jetson and RTX/DGX Spark hardware for the model. Its launch date was announced as 23 October.
Super VL can run on DGX Spark, DGX Station, RTX PRO and RTX systems, NVIDIA Jetson, and in data center and cloud environments. For Bolt 74B-A9B, NVIDIA also lists DGX Spark, DGX Station, GeForce RTX PCs and RTX PRO workstations. The company says the model is also designed for RTX Spark laptops, which are due on 16 October.
Impact on architecture and visualization workflows
For architecture and interior design teams, the announcement’s standout features are Super VL’s ability to process visual documents alongside text, and the Bolt options designed to run on local hardware. The ability to read tables and charts in documents could enable agent workflows involving visual content; however, NVIDIA’s announcement does not promise integration with any specific CAD, BIM or visualization software.
The local execution option may be worth considering for teams that do not want to send data and workloads to the cloud. Hardware compatibility varies by model, though: Bolt 74B-A9B is specified to require 24 GB of memory, while Super VL is available across different NVIDIA systems and cloud environments. Since the source does not explain licensing terms or usage costs, it is worth reviewing access and deployment requirements before moving into production.
Sources
1 sourceSource texts are not republished; short quotes are marked, everything else is our own summary and commentary.
For architecture firms and visualization teams, the practical value of the Nemotron 3.5 family lies in offering options for different scales of work, rather than relying on a single model. Super VL’s ability to read visual documents is noteworthy for document-focused agent workflows, but the announcement does not specify direct integration with BIM or 3D software.
The potential for local use to provide data control is significant, especially for teams working with sensitive project documents. Still, the 24 GB memory requirement, compatible NVIDIA hardware, and licensing and cost terms—which have not yet been disclosed—should all be considered before making a decision. A more cautious approach would be to first test accuracy and hardware demands in a small workflow.
Frequently asked questions
What is NVIDIA Nemotron 3.5 Super VL?
It is an open-weight, 120-billion-parameter multimodal model announced by NVIDIA. The company says it focuses on visual understanding and agent workflows.
When will the Nemotron 3.5 Bolt models be available?
Bolt 74B-A9B is now available. Bolt E4B is scheduled to launch on 23 October.
What hardware does Bolt 74B-A9B run on?
NVIDIA lists GeForce RTX PCs, DGX Spark, DGX Station and RTX PRO workstations. The model was designed to fit in 24 GB of memory.
