In this single‑speaker YouTube review, the presenter evaluates NVIDIA’s newly released free and open large language model, NeMoTron3 Ultra (often mistakenly called NeMoTron3 Super). The video walks through personal hands‑on experiments, noting the model’s astonishing inference speed but also its struggles with complex coding tasks, where it often produces black screens or overly verbose code. After reporting bugs to NVIDIA, some improvements were observed, yet the model remains unsuitable for demanding programming work. The reviewer then highlights the model’s strengths in agentic, system‑administration, and rapid prototyping tasks, emphasizing its utility for everything except challenging code generation. Licensing is examined in depth: the model is released under the OpenMDW license—a machine‑learning‑tailored version of Apache 2.0—earning a near‑perfect rating compared to NVIDIA’s proprietary license. Deployment constraints are discussed, noting the 550 B‑parameter size, 10 % active‑expert mixture‑of‑experts per token, 1 M token context window, and the need for hundreds of GB of GPU memory, leading the reviewer to recommend cloud providers like Lambda. Architectural highlights include mixture‑of‑experts, Mamba layers for efficient long‑range memory, low‑precision NVFP4 arithmetic, and multi‑token heads that predict several tokens in parallel. The model lacks vision or multimodal capabilities, prompting a modular approach of coupling it with separate vision models. The presenter concludes by celebrating the emergence of open‑source competition to OpenAI, praising the open‑science ethos and encouraging the community to adopt and extend such models.
Chapters
This AI is not NeMoTron3 Super. No! This is NeMoTron3 Ultra, NVIDIA's newest free and open AI model. And I've been delighted, disappointed, and confused by it. But I think I got it now. You see, you can look at the benchmarks all you want, but we are Fellow Scholars here. We don't just believe stuff, we test it for ourselves. That is the way of the Scholar. So I had an early look at it and ran some of my experiments day and and night. First impression is that it is incredibly fast. Blazing fast. Love that. But then, my coding experiments did not go that well. When I ask it to write a light simulation program, this is my original area of research, and I get a black screen. Nothing. When I ask it to fix it, it does a bunch of things and— same. Then I said, okay, let's debug this by hand. It had some mistakes. After fixing that, well, we get something. But maybe it's a scene that does not work at all. Other, even smaller systems can do this task with relative ease. And the other thing is, goodness, it wrote up more than 1,000 lines of code. Oof. You don't need that much. My handwritten solution from my research is about 250 lines and renders this scene. Fully open source, free for everyone, forever. Now, let's write a real-time strategy game. Yes! No. Black screen again. Almost. We got a square. But if you ask Deepseek 4 Flash with the same prompt, you get something really cool. But not here. So what is going on here? Well, I went back and forth with NVIDIA and reported some of the issues, and later, there were some improvements. But still, this kind of coding is not something I would personally use this for. So, I said, you know, maybe let's not use this AI. But then I thought, wait, it is super fast and probably good at other things. So, I gave it agentic things. Fixing broken installations on my machine from the terminal? Excellent. Whipping up quick experiments, organizing files? Excellent. Super fast. And over time, I found myself reaching out to it more and more. And I found it to be useful basically for everything other than challenging coding tasks. Now that is excellent, because this might be the openest AI model ever. Weights are open, the research paper on how it was made is open, training data and recipes are being released, at least for the redistributable parts. Now that is pretty crazy. Now, hold on to your papers, Fellow Scholars, because it gets even better. Licensing. Super important question, very overlooked. We are always hoping for Apache 2.0, this is the do whatever you want license. For me, this is 10 out of 10. Now, NVIDIA started publishing their models under their own proprietary license, which I would rate 7 out of 10. Derivative works and commercial use is fine. On the other hand, it needs a bit of attribution and a little stricter on patent grants. Now, this has the OpenMDW license. This is basically Apache 2.0 tailored for machine learning weights. This is absolutely fantastic news. Glorious. I think this might be a 9 out of 10. Maybe as close to 10 out of 10 as you can get from a big company like NVIDIA. Allows basically everything. But less battle-tested, and my understanding is that if you sue claiming this model infringes your rights, you lose the license. Huge improvement. Double thumbs up. Thank you. Now, can you run it yourself? Hmm. Yes and no. Yes, because completely open. Download it, it is yours forever. No limits, no funny business. However, no, because I would love to run it locally too, but it's huge. 550 billion parameters. You need hundreds of gigabytes of GPU memory for that. This is why I will probably use it on Lambda. Also, 1 million token long context window? Ha! Great! Have a larger codebase with a bug hiding somewhere? No worries! Massive books? Easy! Okay, how about images and videos? Well, it does not have vision capabilities. Not multimodal. Text only. Oh man, how much I would love a multimodal version of this goodness. Please! Okay, and I also had a realization. You don't need one model to do everything. You need a roster of models that cover your use cases. For instance, I can't add vision capabilities to Nimatron 3 Ultra, but I can bolt Gemma 4 to it with a screwdriver. It's like a seeing eye dog guiding a smarter blind man along. It is hilarious. And it kind of works. Kind of. So, we finally have more competition in the OpenAI model space, and that is glorious. So, how does it work? Well, one trick is that it is huge, but not all of it runs at once. 550 billion parameters total, but only about 10% of that is active per token. These are specialist mini-brains that are being activated at a time, We call that mixture of experts. But you wise Fellow Scholars know that already, so what else? Now, they also use Mamba layers. Why Mamba? Is this like a snake, or like the fruity chew? I don't know. I don't even know why I brought this up. So what do these do? Well, traditional AI systems have a bit of a memory problem. They work like a student who constantly rereads the textbook over and over again when they are given a question. But memory is precious, so instead, read the book only once and take highly compressed notes. So, this kind of memory remembers important details about the conversation. However, it is also smart enough to throw away the filler words. Thus, this system can process massive amounts of data efficiently. It also uses low-precision numbers, so you have to do less number crunching when running this. They call it NVFP4. And this doesn't rely on predicting tokens one by one. No! It has multiple heads that draft multiple future tokens at the same time. Once again, many things that make it blazing fast. And we get all of this for free, forever. What a time to be alive! Thank you to everyone who worked on this. And absolutely everyone everywhere who is working on open source projects and open models. You are all heroes. And look, this system is great. But it could be tiny, it could be bad, ugly, I don't care. As long as it is open science and open models, it pushes humanity forward. Thank you. What a time to be alive! Here you see me running the full Deepseek AI model through Lambda GPU cloud. 671 billion parameters. Running super fast and super reliably. This is insane. I love it. And I use it on a regular basis. Lambda provides you with powerful NVIDIA GPUs to run your own chatbots and experiments. Seriously, try it out now at lambda.ai/papers or click the link in the description.
