What happened
NVIDIA just shipped Nemotron 3.5 Super VL, an open 120-billion-parameter AI model built for agents that never clock out — and it's not alone. Two edge models came with it: Nemotron 3.5 Bolt 74B-A9B 24GB, tuned to balance intelligence and speed per gigabyte for local agentic workloads, and Nemotron 3.5 Bolt E4B 4GB, tuned for intelligence per gigabyte in fast conversational agents. Super VL and Bolt 74B-A9B are available now. Bolt E4B lands Oct. 23.
That's the news in one breath. The interesting part is what those names actually mean for someone who doesn't work in a data center.
Parameters, edge models, and why you should care
A "parameter" is roughly a dial inside a model's brain. More dials usually mean more capability, but also more memory and more horsepower to run it. So a 120-billion-parameter model is a big brain — and a 4GB model is a small one that still has to be clever enough to be useful.
"Edge" just means the device sitting in front of you: your laptop, your desktop GPU, a small box on your desk. Until now, the smartest agents mostly lived in the cloud. NVIDIA is pushing them down to local hardware. The three open models together cover the whole spread, from large enterprise servers all the way down to edge devices and local machines.
Super VL and Bolt now sit alongside two existing family members. Lightning, released in August, handles high-throughput workloads. Ultra handles the hardest reasoning tasks. Bolt brings intelligence to the edge.
What makes Super VL different
Super VL was built in collaboration with the Nemotron Coalition, and NVIDIA says it delivers leading accuracy on multimodal agentic tasks: following instructions, using tools, running multistep business workflows, and understanding visuals. "Multimodal" just means it can work with more than text — show it a screenshot or a document and it has a way to make sense of it.
The headline number: Super VL gets close to NVIDIA's largest reasoning model while using roughly one-fifth the parameters. On the Artificial Analysis Intelligence Index leaderboard, it reaches up to 5x faster throughput than other models in its class. In practice, that means an agent can grind through a high-volume task — the kind that stretches across days or weeks — without making you sit and wait.
And because it's open, you can post-train it on your own data, run it on your own infrastructure, and keep an eye on what it's doing.
What it means for you
If you've ever wished a chatbot would just handle the boring parts of your day without pinging a server farm, this is that wish moving closer to reality. Here's how it plays out in real life.
At home
Bolt E4B 4GB is the one to watch. It's built for physical and conversational agents on GeForce RTX, Jetson, and RTX/DGX Spark, and it's meant to power fast local voice assistants, interactive characters, and embedded interfaces. Picture a voice assistant that answers instantly because the thinking happens on your own desk — no round trip to the cloud, no awkward pause while it "thinks."
At work
Super VL is aimed at high-volume agentic workflows: instruction following, tool use, multistep business processes, and visual understanding. That's the gap between a chatbot that drafts an email and an agent that opens the right tools, pulls the right data, and works through a checklist on its own. Five times the throughput on the Artificial Analysis Intelligence Index leaderboard means the queue moves faster — which matters when your agent is chewing through hundreds of tasks, not one.
For business
This is where the adoption story gets concrete. Scale AI post-trained Super VL for Skydance's vendor credit finance agent and for EY's M&A due diligence agent. AfterQuery customized it for Hebbia's credit agreement analysis. AI21 and Alice post-trained it for cybersecurity attack resilience. Mercor trained it on the APEX-SWE coding benchmark for coding agents that plan, write, and debug code. ModelScope lets developers fine-tune it through the ms-swift framework for specialized multimodal applications.
Translation: companies aren't just poking at this thing — they're teaching it their own rules, tools, and policies so it gets more accurate on the exact jobs they care about. If you run a business and want a taste of what free AI tools can do before you buy hardware, MyKreaTool is a decent place to start tinkering.
For study
You can hand research, writing, and code debugging to a local agent harness and keep the work on your own machine — no shipping your notes off to a cloud service. For students, that's less about magic and more about not losing your train of thought while a browser tab waits on a server. The agent does the legwork; you do the thinking.
For creativity
Interactive characters and embedded interfaces are the fun corner of this release. A 4GB conversational model that runs locally can drive a character in a game, a kiosk, or an interactive display without a constant internet connection. It's the difference between a puppet on strings and something that responds in the moment.
For income
If you sell automation, the pitch gets easier. You can post-train an open model on a client's data and deploy it on their infrastructure, then charge for the outcome instead of renting someone else's API forever. The local Bolt models mean even a small shop with a single 24GB-class GPU can offer "your data never leaves the building" as a real feature.
How to try it right now
You don't need a research lab. Here's the practical path, free option first.
1. Start free with the open models. Nemotron 3.5 Super VL and Nemotron 3.5 Bolt 74B-A9B 24GB are available now, and they're open — you can post-train them on your own data and deploy them on your own infrastructure. No cloud bill required if you already have the hardware.
2. Check your memory. Bolt 74B-A9B 24GB uses native 2-bit quantization to squeeze into 24GB of memory while running alongside your other apps. Two-bit quantization is a compression trick: it shrinks the model's numbers so the whole thing fits in a smaller box without throwing away the parts that matter.
3. If you're on smaller hardware, wait for Oct. 23. Nemotron 3.5 Bolt E4B 4GB is the lightweight one, built for GeForce RTX, Jetson, and RTX/DGX Spark. That's your on-ramp for voice assistants and conversational agents.
4. Pick your local agent harness. NVIDIA says you can delegate projects to popular local agent harnesses — researching the web, writing, and debugging code — without sending those workloads to the cloud. Use whichever harness you already like.
5. Fine-tune if you need a specialist. ModelScope's ms-swift framework is the named route for fine-tuning Super VL for specialized multimodal applications.
6. Check the scoreboards before you commit. Bolt 74B-A9B 24GB is the #1 open model in its class on PinchBench and the #2 open model for BrowseComp. Those names matter because they measure browsing and agent tasks, not just trivia.
Upsides and what changes
The big shift is that frontier-level intelligence is no longer chained to a giant server farm. Super VL delivers leading accuracy at roughly one-fifth the parameter count of NVIDIA's largest reasoning model, and it runs up to 5x faster than the next-best model in its class on the Artificial Analysis Intelligence Index leaderboard. Bolt 74B-A9B fits into 24GB, which is the kind of memory a serious desktop GPU already has.
For businesses, "open" matters more than benchmarks. You can post-train on your own data, deploy on your own infrastructure, and oversee what you build. No waiting on a vendor's roadmap, no sending sensitive workflows to someone else's servers. And the lineup now spans the whole ladder: Lightning for throughput, Ultra for hard reasoning, Super VL for high-volume multimodal agent work, Bolt for the edge, and Bolt E4B for tiny conversational jobs.
Limitations
Honest version: 120 billion parameters is still a lot of model, and the "local" story comes with a hardware bill attached — Bolt 74B-A9B 24GB needs that 24GB of memory, and "smart enough to get most jobs done" is a real trade-off, not a marketing slip. The smaller Bolt E4B 4GB isn't out until Oct. 23, so if your hardware is modest, you're waiting. The source doesn't spell out licensing terms, pricing, or long-term support, so treat "open" as a starting point for your own due diligence. And no benchmark wins the argument by itself: PinchBench and BrowseComp rankings are useful signals, not proof your specific workflow will work. Running agents all day on your own machine also costs electricity and desk space, which nobody puts in a press release.
Conclusion: One Action for Today
Match the model to the hardware you actually own: a 24GB-class desktop for Bolt 74B-A9B 24GB, or a GeForce RTX or Jetson box for Bolt E4B when it lands Oct. 23. Then grab the open model, fire up your local agent harness, and give it one real task from your week — a research question, a messy document, a small coding bug. You'll learn more in an hour of hands-on use than from any leaderboard.
Comments 0