Nemotron 3.5 Super VL with Native Speculative Decoding

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Nemotron 3.5 Super VL with Native Speculative Decoding · Channel: Fahd Mirza · 11:00 · Watch on YouTube

Topics: AI agents, AI evaluation and benchmarks, Software engineering with AI, Multimodal models

Summary

The video tests Nvidia’s Nemotron 3.5 Super VL with a local app and Hermes agent, starting with a prescription-lens bug that the model successfully finds and fixes, though it also changes files the presenter considers unnecessary. Further tests show strong performance reading Portuguese and handling a hand-drawn circuit, but weaker results in some low-resource language translations and in generating a convincing dragon design from code. The presenter describes the model as fast and capable across several tasks, while noting that its benchmark strengths include long context and speed on Nvidia GPUs. Overall, he considers it a fast all-rounder rather than a clear winner in every category.

Key points

The app bug caused patients with identical eyes but differently written prescriptions to receive different lens calculations and prices.
Nemotron 3.5 Super VL found a missing sphere update in the plus-to-minus prescription conversion and fixed the bug.
After the fix, Daniel’s lab sheet matched Maya’s across measurements and price, but the model also made three changes the presenter thought were unnecessary.
The model correctly read a Portuguese passage from an image and captured its central idea, but some low-resource language titles were poor.
In a hand-drawn Sallen-Key filter test, it read the labels, calculated the cutoff, and noticed that the op-amp inputs were wired incorrectly, with some imprecision in its explanation.
Its single-file HTML tattoo design captured several requested story elements, but the dragon itself was weakly represented.
The presenter describes the model as a 120-billion-parameter model that uses 12 billion parameters at a time, combines Mamba 2, mixture of experts, and attention, and supports up to a million-token context window.
The presenter says the model stands out for long-context performance and speed on Nvidia GPUs, but does not lead all agentic and coding benchmarks.

Chapters

Start	Chapter	Summary
00:01	Prescription app bug and model setup	The presenter demonstrates inconsistent lens results for two patients with identical eyes, then introduces Nemotron 3.5 Super VL and the app environment.
03:41	Bug fix and multilingual image test	The model fixes the prescription conversion bug, but makes extra changes; the presenter then evaluates its reading and translation of a Portuguese image.
06:54	Circuit reasoning and code-generated tattoo	The model analyzes a hand-drawn filter circuit and then produces an HTML tattoo design whose requested elements are only partly successful.
09:28	Benchmarks and overall assessment	The presenter discusses benchmark comparisons, highlighting the model’s speed and long-context performance while calling it a fast all-rounder.

Quotes

It handles up to a million token context window. — 03:08

Daniel now matches Maya on every row. — 04:46

So, this is a real scientific reasoning, not just reading numbers off a page — 07:25

the speed is quite good, roughly double that tokens of its nearly rivals — 09:59

I think it’s a fast all-rounder sort of model — 10:30

Concepts and entities

Nemotron 3.5 Super VL, NVIDIA, Hermes Agent, Lumen RX, Mamba 2, Mixture of Experts, Sallen-Key filter, Fahd Mirza

Full transcript (original language)

00:01 We have a brand new sizzling application to be fixed by a model. This is Lumen RX, an online prescription sunglasses studio. A patient picks a frame and a tint, and the app works out which lens material they need, how thick the lenses will be, whether the frame can take their prescription, the price, and the order sheet that goes to the lab that grinds the lenses. Now, at the moment, the application looks mint, but there is

00:32 a very critical bug in it. Let me first show you. So, here on the top left is a patient, Maya Okonko. I am clicking on her name on the tab, and then in the middle, I’m going to select the frame, which is Apex Sport. And then, I’m clicking on pin to compare one time. Next up, I’m selecting Daniel Reyes. Again, the same frame, and then I’m also clicking on pin to compare. You see, the price has gone up 548.

01:05 I’ll just scroll down to the bottom where a new table would have appeared. There you go. So, we are comparing both of these. So, here is Maya written in minus cylinder on the Apex Sport frame, lens 1.59 6.39 mm thick compatible. Now, look at Daniel. His prescription is written in plus cylinder, but his eyes are identical to Maya’s. Look at the row sent to lab. Now, that is where the things get

01:36 really, really funny. So, what happens is that Maya is minus minus minus, Daniel is also minus, but um the price varies, and both are getting different frames, and there are huge implications to that, as you can probably imagine, because the thing is that two people with identical eyes get different lenses and different prices. A short-sighted customer is up sold to thicker, pricier lenses. A long-sighted customer gets a lens a full

02:06 diopter too weak. So, they pay for glasses that do not fix their vision, and they one of them is even paying more. Nothing is crashing a everything is working, but the whole studio is a mess. That is where I’m going to use this new model from Nvidia, which I will scroll through to show you Nemotron 3.5 Super VL, which is a vision model. And the app is at the moment running locally on my system with a docker, lot of things post grads, front end, back end, redis, as you can see.

02:37 And the model is integrated with Hermes agent. So, I’m going to give this model a goal without telling it anything what needs to be fixed, what is the bug in and all that. So, I’m checking tool use, reasoning, and how soon it is able to fix this bug. As the model works through it, let’s talk a bit more around what Nvidia has done with it. So, this is a 120 billion parameter model that only uses 12 billion at a time. So, it stays fast for its size as per Nvidia.

03:08 The main thing is that it mixes Mamba 2, mixture of experts and attention layers, and it sees as well as reads. Text, images, and video go in, text comes out. It handles up to a million token context window. So, it can hold a whole code base at once, but we will see. It is built for long-running agents that combine reasoning, tool use, and visual understanding, which is exactly what we are testing. It is open weights, which was released just few days back. This is

03:41 Fahad Mirza, and I welcome you to the channel. Please also subscribe to this free weekly AI newsletter at fahadmirza.substack.com. If you’re looking to rent a GPU on very good price, you can find the link to Mass Compute in video’s description. And if you’re feeling too generous, feel free to become a member of the channel. The model has given the response. It has found the real bug as you can see from here, the missing sphere update in the engines plus to minus conversion, fixed it, rebuilt the container, and

04:14 also has verified the result, which is great. But the problem is that it also reported three more fixes in other files, which I don’t think so was needed. Nothing was broken. So, it has just, you know, uh overdone it, which is not good. But anyway, let me go here and refresh this app. I’ll just do control of five. Let’s reload. That’s done. From there, let me select Maya. And I’m going to do the Apex sport. Let’s spin to compare.

04:46 And then I’m going to do Daniel. And from the Daniel Apex, spin to compare. Let’s scroll down. And you can see the comparison that the bug is fixed. Daniel now matches Maya on every row. All the measurements plus the price. So, same frame, same eyes, Daniel’s lab sheet now reads correctly. The bug is fixed. So, the model has done well. And the only three problem is that it has also fixed what was not broken, but anyway, let’s do another test.

05:19 Now, for the next test, I’m giving the model this image of some Portuguese and then asking it to read this paragraph from this image, understand its meaning, and write a three-word title in these 82 languages, including some of the rare ones. So, let’s see if the model is able to translate it, if it is able to capture the essence of this paragraph, which I just showed you, and then if it is able to translate it. So, instruction following multi-linguality, semantics, creativity is there. Let me know once it is done

05:53 because I already have started it or it is already done. It is fast. So, let me show you and then you can give me your opinion in the comments. So, as far as Portuguese image is concerned, it has read it correctly and got the core idea, which is contentment over happiness. And, you [clears throat] know, English, Spanish, French, they all look quite good to me. It followed the format perfectly, one language and title per line, nothing extra. Some of the titles are four words, by

06:23 the way. Anyway, but again, if you look at the, you know, low-resource languages or rare languages, it is failing miserably. Like Burmese is garbage, Malayalam has an English word stuck in it, Pashto, Sindhi, all gone. Several African languages look off. So, I think I would really need some of the native speakers to confirm their languages here, but I think this is an average response because every model these days can go with the

06:54 main European languages and so and so. And let’s do now scientific reasoning with a touch of math. This is a hand-drawn Sallen-Key filter circuit. I’m testing three things here. Can the model read the handwritten values? Can it do the engineering maths? And can it catch that the drawing itself is wired totally wrong? Look at the response of the model. It read every label correctly, identified the circuit, and did the maths cleanly.

07:25 It also has cut off around 750 Hz and few other values, which match my own calculations, by the way. So, all good. Then it did the hard part. It checked the drawing against how the circuit should be wired and caught that the op amp inputs are the wrong way around. So, the filter would not work as drawn. So, this is a real scientific reasoning, not just reading numbers off a page, and it didn’t think long, by the way. So, I think it is quite good. Yes, we can argue that the explanation of the wiring is

07:56 slightly loose because it says there is no output feedback to the minus input, but in the drawing, the output actually loops back to the plus input. So, it is not a perfect answer, but given that the model has been so fast and it has actually got the essence of it, I would call it a pass. So, a good result overall. For the next test, I’m asking the model to design me a dragon tattoo in pure code in one HTML file with a touching American theme, a dragon guarding a

08:27 folded flag and dog tags in memory of a father with a rising sun and a never forgotten banner. No images, just code. I gave this prompt to this model, and you can see that it has generated it after thinking not for too long, and this is what it has generated. I’ll just click on redraw tattoo. So, it is redrawing it. Now, as far as the story pieces are concerned, I think it has got mostly right. The folded flag with the blue

08:57 field and stripes, the dog tags and the glowing sun, the banner with the heart and keyhole, and both view buttons are also there. And the flag and banner is there. Of course, there are a lot of It’s not an image generation software, by the way. Remember that. But there are a lot of weak spots, as you can see. The dragon I don’t even see a dragon. Maybe this is a dragon. No claws, whiskers, or any coiling body whatsoever. So, I would say the theme and layout are okay, but the main subject falls short of a world-class

09:28 stuff, which we would expect from a 120 billion parameter model. If you look at this benchmark table from Nvidia, I will try to make it bit bigger. Look, these benchmarks I really don’t worry too much about them because if you for example look at this. So, against its open rivals, this model is not the overall winner on the agentic and coding test. Quincy 3.8 flash, next GLM 5.2 flash, and deep seek before all

09:59 are ahead of it while it clearly has beaten Quincy 3.5 quite an old model now. But it does shine in long context where it leads the pack. Um the only thing where it is really shining is that on Nvidia GPUs, the speed is quite good, roughly double that tokens of its nearly rivals and we saw it at first hand because uh it really completes the task very very quickly. Now, this chart compares the agent tools themselves and Hermes holds up very well

10:30 as we just tested it out in front of you on terminal batched it land second just behind Codex and ahead of Claude, code, and sweep bench pro. The whole field is packed tightly with Hermes right in the middle of the pack. So, look, I think it’s a fast all-rounder sort of model and overall not a bad model at all. So, let me know your thoughts about this model. Again, please subscribe to this newsletter and if you want to become a member, that

11:00 would be swell. Thank you for all the support.