News update: this post was researched and drafted automatically from the sources listed at the end, and published without a human edit. If you spot a mistake, write to me and I will fix it.
NVIDIA has released three new open models in the Nemotron 3.5 family. Nemotron 3.5 Super VL is a 120-billion-parameter model for always-on agents in the data center. Two Bolt models target local and edge hardware: Bolt 74B-A9B 24GB and Bolt E4B 4GB. Super VL and Bolt 74B-A9B are available now. Bolt E4B comes on October 23.
I pay attention to these releases because my work involves RAG chatbots and agentic coding. Open models that can read documents and run on local hardware matter to clients who worry about data residency and cost. This launch covers three tiers: large servers, desktop workstations, and small edge devices.
In one line: NVIDIA released three open Nemotron 3.5 models: Super VL for large enterprise agents, Bolt 74B for local agentic workloads, and Bolt E4B for fast conversational agents on edge hardware.
What was released
Nemotron 3.5 Super VL is an open 120-billion-parameter model. NVIDIA says it delivers the fastest task completion and highest accuracy for specialized always-on agents. It was built with the Nemotron Coalition. The model handles instruction following, tool use, multistep business workflows, and visual understanding. NVIDIA says its accuracy on these multimodal agentic tasks is comparable with its largest reasoning model at roughly one-fifth the parameter count. It is also 5x faster than the next best open agentic model in its class, according to the blog.
Bolt 74B-A9B 24GB is designed for local agentic workloads on NVIDIA RTX and DGX Spark. It uses native 2-bit quantization to fit in 24GB of memory. NVIDIA calls it a balance of intelligence and speed per GB. Bolt E4B 4GB targets GeForce RTX, Jetson, and RTX/DGX Spark for physical and conversational agents.
Why Super VL matters for document-heavy work
Many client agents still treat text and images as separate problems. Nemotron 3.5 Super VL combines text and visual understanding in one model. NVIDIA says it processes a web page, extracts text, reads graphs and charts, and returns insights. The same applies to PDFs and other documents.
This matters for RAG chatbots. A contract or invoice often contains tables, logos, stamps, and handwritten notes. A text-only pipeline misses these. With Super VL, the model sees the page and acts on it. NVIDIA says agents do not need to route tasks between separate systems. Context carries through as work moves from document to data to decision.
The model is open. Open weights, open datasets, and training recipes are available. Organizations can post-train Super VL on their own data and run it on their own infrastructure. NVIDIA reports adoption from Scale AI for Skydance and EY, Hebbia, Atlassian, CrowdStrike, Perplexity, Salesforce, and others. I find the insurance and financial services examples most relevant: a customized agent can scan claims for fraud markers or review credit scores against a lender's criteria on premises.
Bolt brings agentic work to local hardware
Bolt 74B-A9B 24GB is the model I am most likely to test first. It runs locally on RTX or DGX Spark, so workloads do not need to go to the cloud. NVIDIA says users can delegate projects to local agent harnesses for web research, writing, and debugging code. It is the number one open model in its class on PinchBench and number two open model for BrowseComp. I cannot verify those benchmarks from this single source, but the design goal is clear: enough intelligence to get most jobs done, faster than the smartest models.
Bolt E4B 4GB is smaller and meant for fast conversational agents. It targets voice assistants, interactive characters, and embedded interfaces. The October 23 release date means I will wait before assessing it.
What these models do not yet solve
NVIDIA says Super VL is comparable with its largest reasoning model on multimodal agentic tasks, but the Nemotron family still has Ultra for the hardest reasoning tasks. So a team with very hard reasoning requirements may still need Ultra or a different model. Super VL is also 120 billion parameters. It is efficient relative to larger models, but it still needs capable infrastructure. The blog does not claim it runs on a laptop. For local use, Bolt is the option, and Bolt E4B is not available yet.
Openness does not remove integration work. Fine-tuning with NeMo and open datasets requires effort. The source lists many partners who have already done this, but a small team like mine would start with a pilot, not a full rollout.
What I would try
For a client with confidential financial or medical documents, I would run a small evaluation of Super VL against the current text extraction plus LLM pipeline. I would test OCR on scanned tables, chart reading, and multi-step questions that require the model to see a document and then call a tool. If accuracy is good enough, I would consider fine-tuning with NeMo on the client's own document types.
For my own agentic coding setup, Bolt 74B-A9B 24GB is worth watching. Running a local model for code fixes removes cloud upload concerns. I would need a workstation with 24GB of GPU memory. The source says it runs alongside other applications, which matters for a working machine.
I am not changing production systems today. The release is interesting because it covers the full range from data center to edge, but I need to see how Super VL performs on Indian client documents, especially PDFs with mixed English and regional language text. The source does not mention multilingual performance, so that remains a gap to test.