NVIDIA Nemotron 3.5 Guide: Pick and Run Open Agent Models

Last updated: October 16, 2026

NVIDIA quietly dropped one of the most practical open-AI releases of the year on 15 October 2026: three new open-weight models built specifically for AI agents. The headline is Nemotron 3.5 Super VL, a 120-billion-parameter multimodal model aimed at always-on enterprise agents — and the sleeper hit is Nemotron 3.5 Bolt 74B-A9B, a 74-billion-parameter model squeezed by 2-bit quantization into just 24 GB of memory, so it runs on a single high-end RTX workstation. A third model, the 4 GB Bolt E4B, follows on 23 October.

Quick answer: NVIDIA's 15 Oct 2026 open release: Super VL (120B, 1M context), Bolt 74B (24GB, RTX 5090) and Bolt E4B (4GB edge, 23 Oct) — all free commercial-use weights.

All three ship with open weights under a licence that allows commercial use, plus training recipes for anyone who wants to adapt them to their own data. If you run agents, evaluate models, or just want to know which open model is worth your hardware, this guide maps the release to real decisions: which model fits which machine, what the licence actually permits, and how to get started today.

What NVIDIA actually released on 15 October

Three models, one family, three deployment tiers — confirmed across multiple reports published the same day (oossa, unite.ai, wccftech, maccus.fr, all 15 Oct 2026):

Nemotron 3.5 Super VL — 120 billion parameters (12 billion active per token, mixture-of-experts), 1 million token context, multimodal (text, images, short videos, PDFs). Built for specialized always-on agents: reading documents, checking information across sources, following company policy, calling several tools in one request. Available now.
Nemotron 3.5 Bolt 74B-A9B 24GB — 74 billion parameters (9 billion active per token), 256K context, native 2-bit quantization. Runs locally on RTX PCs and workstations — NVIDIA tested it on a PC with a GeForce RTX 5090 and on DGX Spark. Intended for local agent workloads: web research, writing, code debugging without sending anything to the cloud. Available now.
Nemotron 3.5 Bolt E4B 4GB — compact conversational-agent model for GeForce RTX, Jetson and RTX/DGX Spark systems. Arrives 23 October 2026.

The release also includes open datasets and training recipes, and NVIDIA says Super VL was built with the Nemotron Coalition. In NVIDIA’s lineup these sit alongside Lightning (high-throughput workloads, released in August) and Ultra (the hardest reasoning tasks), with Bolt covering the edge.

Which model fits you: the decision table

Open models are only useful if you can actually run them. This is the desk-built decision matrix from the verified specs — read the row that matches your hardware:

Model	Params (active)	Context	Hardware NVIDIA names	Best for	Out?
Super VL 120B-A12B	120B (12B)	1M tokens	1 DGX Spark, 1 H200, 2 H100, or 1 GB200	Always-on enterprise agents, document-heavy multistep work	Now
Bolt 74B-A9B 24GB	74B (9B)	256K tokens	RTX PCs (RTX 5090 tested), DGX Spark, RTX Spark laptops	Local agents on a workstation: research, writing, code debug	Now
Bolt E4B 4GB	Not disclosed	Not disclosed	GeForce RTX, Jetson, RTX/DGX Spark	Fast conversational agents on tiny devices	23 Oct

The honest headline: Super VL is server-class hardware territory. If you have one serious workstation with 24 GB of VRAM, Bolt 74B is the model built for you. And if you are waiting for an edge device, E4B arrives next week.

The mixture-of-experts maths that changes the hardware question

Both launched models are mixture-of-experts designs: only a fraction of the parameters activate per token — 12 of 120 billion in Super VL, 9 of 74 billion in Bolt. Compute per token is roughly that of a 12B or 9B model, which is why Bolt can be useful on a workstation while the full 74B still has to fit in memory (the 2-bit quantization does that part).

NVIDIA’s performance claims, all labelled as the company states them:

Super VL delivers up to 5x the throughput of the next best open model in its class for agentic tasks — NVIDIA points to the Artificial Analysis Intelligence Index leaderboard for this figure. It is a company-cited number, not an independent result.
Super VL reaches accuracy “comparable to NVIDIA’s largest reasoning model at approximately one-fifth the parameter count” — again, NVIDIA’s own claim, on specialized agentic tasks.
Independent-tester figures: tech analyst Ryan Shrout posted on 15 October that in his testing, the new Super build finishes 4x the share of enterprise jobs its predecessor did and spends 123x fewer tokens per correct task. Treat this as one tester’s result, not a guarantee for your setup.

The takeaway that matters: for agentic work, tokens-per-correct-task is the metric to watch, not just raw parameter count — and both these models are engineered to be efficient there.

Can you use it commercially? The licence in plain English

Both launched models ship under the OpenMDW-1.1 licence, which allows commercial use (confirmed by maccus.fr, 15 Oct 2026). In practice:

You can: download the weights, run them in your products, post-train them on your own proprietary data, tools and workflows, and deploy on your own infrastructure for privacy and control.
Enterprises already are: NVIDIA says dozens of companies have fine-tuned Super VL on private data, naming Scale AI (finance agents), AfterQuery (credit-agreement analysis) and TrainLoop (dermatology prescription drafting); platform partners include Atlassian, CrowdStrike, Salesforce and Perplexity (oossa, 15 Oct 2026).
The watch-outs: the release is built around NVIDIA hardware (the 24 GB figure was tested on a GeForce RTX 5090); and while datasets and recipes are open, post-training a 120B-class model well still needs serious compute and expertise.

How to get started: 5 steps

Desk-researched on 16 Oct 2026 — this is the route from the announcement, not a hands-on test:

Find the weights. The models are published on Hugging Face, and also listed on OpenRouter and ModelScope for hosted access (oossa, 15 Oct 2026). Search the NVIDIA org page for “Nemotron 3.5” — do not download from random mirrors.
Check your hardware against the table above. Bolt 74B needs 24 GB of GPU memory; Super VL needs server-class hardware (1 DGX Spark, 1 H200, 2 H100 or 1 GB200, per NVIDIA). Match the row before you spend download time.
Pick an inference stack. For MoE models this size, the standard open routes are vLLM, SGLang or Hugging Face Transformers — check the model card for the recommended configuration for each model before committing.
Or try hosted first. OpenRouter and ModelScope availability means you can test Super VL or Bolt on real tasks before buying or freeing up hardware.
Then decide: fine-tune or stay stock. NVIDIA ships training recipes, so a company can adapt the model to its own data and deploy privately — but if you don’t have labelled agent data and post-training experience, the stock model plus good prompts is the sane starting point.

Our take: the decision matrix that actually matters

Here’s the honest framing nobody in the launch coverage gives you: these models are not competing with ChatGPT or Claude on your phone. They are competing for the slot between closed APIs and giant open models — the “I want a serious agent model I control” slot. Super VL is for teams with servers; Bolt 74B is for the solo builder with one good GPU; E4B is for embedded conversational agents.

Related reading on this site: our Step 5 open-weights download and run guide (another fresh open release with a commercial licence), our OpenCode 2.0 setup guide (an open coding agent that benefits from strong local models), and our x401 agent-identity explainer — because the more agents you run locally, the sooner you need to know how to trust them.

Frequently Asked Questions

Can I use NVIDIA Nemotron 3.5 models commercially?

Yes. They ship under the OpenMDW-1.1 licence, which allows commercial use — including running them in products and post-training on your own data. Verify the licence text on the Hugging Face model card before production use.

How much GPU memory do I need for each Nemotron 3.5 model?

Per NVIDIA: Bolt 74B-A9B needs 24 GB (tested on a GeForce RTX 5090 class card); Super VL 120B needs server-class hardware — one DGX Spark, one H200, two H100s, or one GB200. Bolt E4B’s specs are not yet fully disclosed; it arrives 23 October 2026.

When is Nemotron 3.5 Bolt E4B (4GB) available?

23 October 2026, per NVIDIA’s announcement. It targets GeForce RTX, Jetson and RTX/DGX Spark for fast conversational agents on small devices.

Do I need NVIDIA hardware to run these models?

Effectively, yes, for now. NVIDIA’s test and reference hardware is all NVIDIA (RTX 5090, DGX Spark, H100/H200/GB200), and the models are optimized for its stack. Cross-platform support is not part of the announcement.

How does Super VL compare to closed models for agents?

NVIDIA claims Super VL matches the accuracy of its largest reasoning model at roughly one-fifth the parameters on specialized agentic tasks — a company-reported claim, not an independent leaderboard result. Independent testing (e.g. Artificial Analysis figures NVIDIA cites) is the metric to watch before betting on it.

Sources and methodology

All launch facts (dates, specs, hardware requirements, licence, availability) are from reports dated 15–16 Oct 2026: unite.ai, wccftech.com, maccus.fr and oossa.com. Performance figures are labelled per source: NVIDIA company claims vs independent-tester figures. The decision table and readiness framing are original desk analysis; the models were not run hands-on for this article. Verification date: 16 October 2026.

Govind Dheda

Govind Dheda covers AI tools, guides and prompts for OpenAIMaster — what is new, what is worth using, and how to put AI to work.

Have a burning question about this topic?
Feel free to email us at contact@openaimaster.ai — we are happy to help!

NVIDIA Nemotron 3.5: How to Pick the Right Open-Weight Agent Model and Run It Yourself