Nemotron 3.5 Super VL: FREE AI for Hermes Agent!

Julian Goldie (0:00)

So, NVIDIA just released a free, open AI model. It can read your screenshots, your PDFs, your charts, and it can actually go and do the work from there. It's called Nemotron 3.5 Super VL, and it just dropped. As you're on a test built for Hermes Agent, it'd be NVIDIA's own bigger model, Nemotron 3 Ultra, and you might be thinking, okay, why should I care about this? So, a lot of the work that we do inside of this is very visual. And with this system, it can basically read documents, it can interpret charts, it can search the web to complete a task. And it's also very efficient for coding, plus you can actually use it for free, and I'll show you how in a second. So, if we go to our Gentic operating system here, and we go to Open Code, you can use it for free inside Open Code in the drop down here. So, we can select Nemotron 3.5 Super VL, then we can say, for example, just build a beautiful to-do list, make it entertaining and engaging, et cetera. And then we can ask it to build over there.

Now, we can also use this inside Hermes Agent. So, that means you got free models that you can use inside Hermes Agent, as you can see right here, and also inside Open Code. So, today I'm going to walk you through what is, what's new, what's the numbers actually saying, and where it fits into your setup. Let's get straight into it. And let's talk about the three numbers you probably want to be aware of right here. So, the first is 120 billion. That's how big the model is in total, but only 12 billion in those parts. Switch on for each word it writes. That's why it's pretty fast. Second is up to five times. So, NVIDIA actually says on one graphics chip, it gets through agent work up to five times faster than other open models in size. And third, 63.44%.

That's its score on Hermes Bench. So, NVIDIA's bigger ultra model scored 56.33%.

And this is a multi-modal model as well. Now, a lot of agents today just run on text models. So, obviously, when you hand a text model to a picture or something like that, it just can't use it, right? Whereas a vision model like Nemotron 3.5 Super VL is more agentic because agents might need to read documents, look at charts, search away for one task. And this model puts text and visual understanding in one model, so you don't pass context between a separate text model and a separate vision model. And it's improved a lot on performance as well versus previous models here. You can actually see, for example, versus Gemma 4

And you might be wondering, okay, why is it being compared against Gemma 4? Because that's Google's open model, and you can see how it performs on the benchmarks right here. Now, there's also a test called CharXIV, right? And it checks if an AI can really understand charts. Now, this model actually scored 80.8%.

NVIDIA's smaller model from here before, Nemotron 3 Nano Omni, scored 63%.

And when it's allowed to run a bit of code to help, it actually hits 84% on performance. It's also near the top of the open models on a test called OCR Bench V2 OCR. And OCR just means reading words and sight pictures, like reading the text on a receipt photo, for example. Now, the second new thing is speed. So if we go back to our build here, we've open code. You can see that the build is already ready to go, and that works nicely. Again, this is not like a frontier model. I just want to be clear here. So this is not like, you know, Opus 5 and 5 or something like that. This is a new open source model that's free to use faster. And you can use it inside your agents. That's the difference. That's why you want to think about this one properly. Now, with speed, it also has a small footprint. So if you remember, only 12 billion parts switch on at a time. If you picture a huge team of like 120 people for each job, only the 12 who are best at it stand up. Everyone else stays seated.

So you basically get the brains of a big model with the speed of a small one. And on one test with a really long input of 100,000 tokens, it wrote over 600 tokens per second, which is really fast. So that's about 100 words second, whilst holding a huge amount of information in its head. NVIDIA also made it run on DGX Spark. So you can actually run it on DGX Spark. That's basically like a small NVIDIA computer that sits on a desk and a group called EXO Labs tested on that machine. They actually found it gave one of the best mixes of right answers and speed for agentic work out of the open models that fit onto it. And the third new thing is that it's open. So NVIDIA have released the model weights, which is the brain itself. They released the training data, they released the recipes for how they trained it, and they put out 40 vision datasets too. That means you can fine tune it, you can give it your way of working. If you have a DGX Spark or you can just use it for free on something like OpenCode and you get to go, as you can see right here. There are cool things like you can switch between models. So for example, if you run out of tokens on Nemotron 3 Super or you need something else, then you could switch to Cat Jam or Big Pickle with OpenCode, or you could use it gentrically inside Hermes Asian as well. You got a lot of choice when using this model. So you might be wondering, okay, well, where can you get it? It's already in a few places. It's free right now inside OpenCode. It's not an open router, which lets you plug one model into lots of apps. You can get it from Hugging Face, Olama, and also LM Studio. Supports if you want to run it on your own computer. Again, you would need something like a DGX Spark to run this. And you could also use it inside Hermes Agent and OpenClaw as well. Now inside my Gentic operating system, like you see right here, we've got Hermes, we've got OpenClaw, we've got OpenCode, we have Claude, all working side by side inside a powerful Gentic operating system, all sharing one model. So a model like this is something you can just slot in for the job that starts with a picture or file. And the same goes for whatever agent you use. You don't need my setup. Any agent that lets you pick a model for OpenRouter can try it. Now, if you want help actually sending us help, this is what we actually do inside the AR Profit Boardroom. Link in the comments description or go to the arprofitborm.com so you get our Gentic operating system. You can plug in Claude, Hermes, OpenClaw, and you can swap in new open models, for example, like Nemotron 3.5 SuperVL for the jobs, where your agent needs to read screenshots, client reports, or receipts. I've also added tools inside the Gentic OS for SEO, for avatars, for videos, and you get the zip file, a video tutorial, a 30-day roadmap to set up and updates as we improve it. Plus, you get four weekly coaching calls where you can ask questions, get help and support in real time, and learn how to use new models like this to save time and get more customers. Link in the comments description or just go to the arprofitborm.com to get access. Now, you also might wonder how it performs on head-to-head tests. So, Hermes Bench Nemotron 3.5 SuperVL scored 63.44%, Nemotron 3 Ultra scored 56.33%.