NVIDIA released Nemotron 3.5 Super VL on October 15, 2026, an open 120-billion-parameter model it presents as the fastest and most accurate in its class for always-on agents. Two smaller Bolt models ship with it, one sized for a 24GB workstation and one for 4GB devices. Super VL and Bolt 74B-A9B are available now, and Bolt E4B arrives October 23.
What was actually released
Super VL is a vision language model for what the post calls always-on agents: instruction following, tool use, multistep business workflows and visual understanding. Its vision work covers optical character recognition, tables, charts and structured information in documents, so an agent reads a page, a PDF or a chart and acts on it instead of handing the task to a second model.
The release carries more than weights. NVIDIA says the model ships with open datasets and training recipes, that customization runs through NeMo and Python pipelines, and that Hugging Face gets datasets for tool use, finance, instruction following, long context, fairness and vision-language understanding, plus reinforcement learning environments in NeMo Gym. Super VL is available now on Hugging Face, Open Router, ModelScope community and as NVIDIA NIM, refreshed every two weeks with security patches. It runs across DGX Spark, DGX Station, RTX PRO and RTX at the desktop, on Jetson at the edge, and in the cloud.
Bolt 74B-A9B uses native 2-bit quantization to fit inside 24GB of memory and run next to other applications, and is on Hugging Face now with support for llama.cpp, Ollama, LM Studio, Unsloth and vLLM, ahead of the RTX Spark laptops NVIDIA says arrive October 16. Bolt E4B is the 4GB model for GeForce RTX, Jetson and RTX/DGX Spark. Both join Lightning, shipped in August, and Ultra.
| Model | Footprint | Where it runs | Availability |
|---|---|---|---|
| Nemotron 3.5 Super VL | 120B parameters | Data center to desktop | Now: Hugging Face, Open Router, NIM |
| Nemotron 3.5 Bolt 74B-A9B | 74B, fits 24GB at 2-bit | RTX Spark laptops, GeForce RTX, DGX Spark | Now, on Hugging Face |
| Nemotron 3.5 Bolt E4B | 4GB | GeForce RTX, Jetson, RTX/DGX Spark | October 23 |
What ships on day one
- Super VL: open weights, datasets and training recipes, tuned through NeMo.
- Bolt 74B-A9B: a 24GB footprint, the number that picks the workstation.
- Bolt E4B: promised for October 23, not today.
What the claim rests on
The performance line is a throughput comparison, not a benchmark table. NVIDIA writes that Super VL reaches up to 5x faster throughput than other models in its class, on the Artificial Analysis Intelligence Index leaderboard, and the class is never defined by parameters, hardware or prompt length. The two rankings the post does name, first on PinchBench and second on BrowseComp for Bolt 74B-A9B, arrive without the scores behind them.
The accuracy claim has the same shape. Super VL delivers, in the post’s words, the fastest task completion and the highest accuracy for specialized always-on agents, and NVIDIA says it is comparable with its largest reasoning model at roughly one-fifth the parameter count. Read that as arithmetic: at 120 billion parameters, the comparison model sits near 600 billion, which is our calculation and not a figure NVIDIA publishes. The one-fifth claim is the only one a buyer can test on hardware they already own. Nothing in the post says what the throughput numbers were measured on: no context window, no latency, no tokens per second, no cost per task.
The training figures are inputs rather than results: about 26 trillion tokens across pretraining and continued pretraining, then post-training on roughly 3.8 million samples with reinforcement learning across dozens of environments. The adoption list is a customer list. Scale AI post-trained the model for a credit finance agent at Skydance and for M&A due diligence at EY, AfterQuery for credit agreement analysis at Hebbia, AI21 and Alice for cybersecurity. Atlassian, CrowdStrike, Dassault Systemes, Perplexity, Salesforce and Synopsys are named as platforms integrating it. None of them publishes an evaluation of the model in the post.
| What the post claims | What backs it |
|---|---|
| Up to 5x faster throughput | A third-party leaderboard named, the comparison class undefined, no configuration |
| One-fifth the size of the largest reasoning model | No model named, no benchmark, no parameter count for the comparison |
| Highest accuracy for always-on agents | No accuracy table, no task list with scores |
| First on PinchBench, second on BrowseComp | Rankings without the numbers |
What the announcement does not carry
No price and no license terms for any of the three models. No context window, no latency, no throughput in tokens per second, no independent evaluation. The Dynamo line, 8x more users at 85% lower latency, has no hardware configuration and no baseline.
On this page
What it changes for teams choosing an open model
For a buyer, the interesting number is 24GB. It is the point where an agentic model stops being something rented per token and becomes something a workstation can hold. Bolt 74B-A9B is aimed straight at that line, and the harness list, llama.cpp, Ollama, LM Studio, Unsloth and vLLM, is the practical half of the announcement: the model drops into the tooling people already run.
The second change is what leaves the vendor’s hands. Open weights plus datasets plus recipes plus a post-training path mean a team can tune the model on its own documents and keep the result. That is the trade we described in Google Cloud’s Gemini agent with hard spend caps: control is what closes the sale, and control over an installed model is stronger than a limit set by a provider. The cost moves with it: on your own machines the bill becomes hardware, electricity and engineering time, the part the post leaves unpriced.
We covered NVIDIA’s other side yesterday in its European city deployments. That post sold computing to cities through partners and carried no price either. Two announcements in two days, two missing price lists, same vendor.
The Neural Minute take
The headline is speed. The argument is size, and only one of the two can be checked without NVIDIA’s help.
A 5x throughput claim against an undefined class says nothing about a buyer’s own workload: the same model on a laptop and on a rack of accelerators produces different numbers. Parameters, quantization and memory fit can be measured on hardware a team already owns, and 24GB is a number that survives contact with a real machine.
Any organization that wants to run agents on NVIDIA infrastructure now has an open, optimized model for their tier, without having to choose between capability and deployability.
That sentence is the commercial claim in one line: NVIDIA is selling a ladder, and the ladder is NVIDIA hardware. If a tier is defined by memory, the model choice follows the machine a team already bought. What the post does not answer is what happens on a machine that is not NVIDIA’s, and for a site that covers tools people pay for, that decides whether this release matters beyond the vendor’s own customers. Our rule from our report on agents that overstate their results still holds: a number that arrives without a method is a direction, not a measurement.
What would change our read: a benchmark table with hardware and prompt length attached, an independent evaluation of the speed claim, a published price for NIM, or results from a team running Bolt 74B-A9B on a 24GB card it owns.
Sources
- NVIDIA, the post cited here, published October 15, 2026: https://blogs.nvidia.com/blog/nemotron-3-5-super-vl/
- Our report on NVIDIA’s European city deployments, read on October 15, 2026: https://www.neuralminute.com/news/european-cities-deploy-nvidia-ai-agents/
- Our report on a Gemini agent shipped with spend caps, read on October 15, 2026: https://www.neuralminute.com/news/google-cloud-gemini-agent/