NVIDIA'S Small Model: Nemotron-3.5 Bolt E4B 4GB Locally

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: NVIDIA’S Small Model: Nemotron-3.5 Bolt E4B 4GB Locally · Channel: Fahd Mirza · 10:50 · Watch on YouTube

Topics: AI evaluation and benchmarks, Software engineering with AI, AI models, Local and open-source models, Multimodal models

Summary

The video tests NVIDIA’s small Nemotron model locally, using a race-control app bug and several other tasks. In the coding tests, the model fails to identify the bug and produces an HTML scene that does not match the request, while incorrectly claiming the bug fix was complete. It performs better on a trick riddle, a basic physics problem, and a scheduling puzzle, though its multilingual name suggestions contain repetitions and generic choices. The creator’s overall view is that it is a decent model for its size, but coding is not its strength.

Key points

Nemotron is described as a post-trained model based on Google’s Gemma 4 E4B, with 8 billion total parameters and 4 billion active parameters.
The tested Q4KM version uses about 6 GB, while the local run shown uses just over 4 GB of VRAM.
The model failed to find the race app’s effort-measurement bug and wrongly reported success.
Its generated HTML scene lacked most of the requested yakisoba details and visual elements.
It correctly identified that a rooster cannot lay eggs, but did not question the rest of the riddle’s account.
It gave a good answer to the ball-drop math and physics question, showing its steps.
It performed well on a staff scheduling puzzle, with a couple of slips noted.
The multilingual test followed the requested order for 82 languages, but included repeated and generic names.
The creator considers the model reasonable for its size and potentially useful for everyday tasks, but not strong at coding.

Chapters

Start	Chapter	Summary
00:01	The race-control bug and local model setup	The creator introduces a fictional ultra-marathon app bug and runs Nemotron locally through llama.cpp with a Hermes agent.
02:12	Model details and benchmark claims	The video describes the model’s size, memory use, intended strengths, and NVIDIA’s benchmark positioning.
04:18	Coding tests	The model misses the app bug and falsely claims success, then generates an HTML scene that falls short of the requested description.
07:03	Riddle, creative writing, and physics	The model spots the rooster riddle’s obvious trick and gives a solid step-by-step answer to a basic physics problem.
09:10	Scheduling, multilingual test, and verdict	The model does well on a scheduling puzzle but produces repeated or generic names in the 82-language test, leading to a mixed overall verdict.

Quotes

The app counts the downhill meters instead of the uphill ones — 01:05

it has even declared victory. — 04:52

So, it passes the riddle — 08:07

coding doesn’t really seem a forte of this model — 08:39

not a bad model at all from Nvidia given its size — 10:50

Concepts and entities

Nemotron 3.5 Bolt E4B, NVIDIA, Fahd Mirza, llama.cpp, Hermes, Gemma 4 E4B, Corima Live, Rarámuri

Full transcript (original language)

00:01 Meet the Rarámuri tribe in Mexico whose name means the foot runners. They cross the Copper Canyon on foot, often in simple sandals, and running is part of their culture, not a sport. And this is Corima Live, race control for the Barrancas 80K, a fictional 80-km run from the Canyon Ultra Marathon inspired by this Rarámuri tradition. It tracks 36 runners and forecasts when

00:34 each one finishes, who will miss the cut off, and how long the medics and sweep crew must wait at the finish line. Right now, the application is running fine, all clear, nobody will miss the cut off at all, and the crew can go go home at 19.18. But, the thing is that this app has a bug. The bug is that how it measures effort. Look at these boxes here. Now,

01:05 if you look at the very last one, that is the actual bug. This leg has a 760-m climb, but shows a low effort number, and this one is almost flat, but shows a high one. The app counts the downhill meters instead of the uphill ones, so the canyon looks easier than it is, and every forecast comes out too early. So, let’s see if this a new model from Nvidia can fix this or not. This is a small 4-billion effective parameter model built for fast

01:38 local assistant for your local systems or edge devices, and I’m currently running it on my local system. This is llama.cpp, which is serving the model, and this is my VRAM consumption at the moment, which is just over 4 gig of VRAM in this context of format which Nvidia has released. Now, I already have configured it with my Hermes agent where I’m going to give it this goal to go and find this bug in app and then tell me how exactly

02:12 to run this again. So, let’s see what it does. While that happens, let me take you back to the benchmark and we will also check out the model detail. And by the way, this is the application which is currently running on my local system, a full stack application with lot of, you know, tiers around back end, front end, and lot of Redis, too. Now, this is Fahad Mirza, by the way, and I welcome you to the channel. Should have said said it earlier, but anyway. Also, I have recently started this free

02:43 weekly AI newsletter which you can subscribe at fahadmirza.substack.com. So, let’s go back and talk bit more around this model as what exactly is this. So, as you can see that it is a small language model which is post-trained from Google’s Gemma 4 E4B. It has got 8 billion total parameters of which 4 billions are active. And the Q4KM versions we are running is only about 6 gig. It is built for fast local voice and device assistant. So, it is tuned for low memory use and quick first

03:16 token with a focus on multi-turn tool use, instruction following, and sticking to the point uh where you input is fast through and then it just supports 64K tokens of context quite easily. If you look at this benchmarking image from Nvidia, the idea is very simple. Higher up means smarter, further left means faster to respond. Bold lands in the green smart and fast corner, which means that it is beating the other small models around it on score including the

03:47 Gemma 4 model it was built from while responding about as fast as the smallest one. Only one model on the chart scores higher, which is Gemma 4 26 billion A4B, but of course it is a much bigger circle, so it uses far more memory and responds slower. And if you look at the big one on the far right, that is the slowest of all and still scores lower. So, what Nvidia is trying to say here is that this is a small model that punches above its weight. So, let’s see if that

04:18 holds up on this real bug or not. Let’s go back to our Hermes agent. It is still working through it, so let’s wait. The model has come back with the response. And if I look at the result and also thinking from this, it has missed it completely. The model uh read the files but never really opened anything, uh especially the file where it the bug lives. It also planned to read some files but never did. So, it only checked that the color responses had data, never whether the

04:52 values were right. It never even compared an effort number to the formula. So, even if I just go here and I will just do control of five in front of you. So, let’s scroll down and you will see that nothing has changed. The bug is still there. But if you look at the response, it has even declared victory. I’ll just go down and show you. There you go. You see? So, it was installing something, then it just moved along. And the you know, it says that a goal

05:25 has been achieved, which is totally wrong. So, I’m just giving you the honest straight picture out of this model. Now, I have used Gemma 4. You can check out that video on the channel even in contest format, quite an old model, plus the coin one. Both of them never did this thing where they just hallucinated and gave us the false response. Anyway, let’s do another test before we make any judgment. Okay, next one was where I asked the model to create me a single HTML file

05:59 where a steel teppan griddle seen from above, a pile of yakisoba noodles, and then basically a noodle is being cooked, mixed in, and then browning slowly. So, it is not a hard one because I understood that it is a small model, so let’s keep it simple. And then it has thought a bit. After thinking, uh you can see that the thinking is not too deep and profound, which is also acceptable. And then it says that it has written the file and verified it, too. And this is what it has created. So, you see there

06:30 is no pile of noodles, no cabbage, no carrot, no steam at all, and no texture at all, just a flat gray background with a few white dots, two gray sort of sticks standing in for spatulas, and one dark red curve for the sauce, I guess. Um so, the model has delivered a file, but not the scene we asked for. So, you can make your guess, too, if this test is a pass or fail. Okay, now let’s me do a trick riddle. So, you can see I’m asking it that a

07:03 rooster uh you sorry, I have 11 chickens in my yard. For the last few days, none of them was laying any egg. And I became quite upset today. Angrily, I went into their den, and all of them got scared and laid two eggs each, except one who just laid one. I was surprised and asked that chicken, “Why did you lay only one? Are you not scared of me?” Chicken said, “Sorry, I am a rooster, and that’s all I could do.” My question is, how is that possible? Plus, write a creative paragraph in Mark Twain’s style.

07:34 So, of course, I mean, I’m just checking if this model knows that a rooster cannot lay an egg. And then, um you see that model has risen a bit. So, basically I’m testing if it can follow the logic plus creative writing. The model spotted the trick that a rooster cannot lay egg. And then you see that it has given me this answer. Sorry, I’ll just go here. There you go. So, this is a believable mark to an voice. I can say that. It stopped at the obvious catch and never questioned the account.

08:07 So, it passes the riddle, but I think it couldn’t do the deeper check, which is expected of this model. So, this test is a pass. Creative writing seems good, which means that for day-to-day use cases, general answer question on your mobile device, I think this is not a bad choice at all. Next, we are testing scientific reasoning with math. We dropped a 2 kg ball from a 45 m building and asked the model for the fall time, the impact speed, and the impact energy. And then you see that uh we are also making it to

08:39 prove its own answer. And look at the result. 3 seconds, 30 m/s, and 900 J. Every step shown, and the energy check matches at 900 on both hands. For a 4 billion model, I think this is a very good answer for scientific reasoning and math. So, coding doesn’t really seem a forte of this model, but scientific reasoning, math, not a very high-fi, but a decent one. Plus, some of the creative general writing is okay.

09:10 And now I am doing a rostering test. So, you see this is basically a scheduling puzzle that staff two events every Saturday, one needing a French-speaking waiter and one a Hindi-speaking waiter, using each chef’s and waiter’s availability while avoiding two chefs and waiter pairs who don’t get along. The model worked out that only Luca speaks both languages and put him with Chef Liam on the 15th and 29th and correctly ruled out the first, eighth, and 22nd. And I think model has done wonderfully

09:43 well here. There were a couple of slips where a model note was given, but overall it’s a clear pass. That’s a good one. Finally, let me do a multilingual test. So, I’ll just run this. So, what I’m asking the model here, generate the best possible name of baby boy in all following languages, and these are all the languages of the world. So, let’s see if it is able to do that or not. The model has at least followed the instruction, as you can see. It counted all 82 languages in the same order. But if you look at the response,

10:18 there is a varied response. Lot of slip is here. Lot of repetition of name. Um and then for example, for Urdu Saraiki, same name is being repeated. Same for Malayalam and Nepali, Italian and Esperanto. So, the training set doesn’t you know, seem too profound. Bjorn for Icelandic and Gutnish. So, some picks are quite generic rather than language-specific. But let me know your thoughts too after checking out your languages. So, look, I think not a bad model at all

10:50 from Nvidia given its size and the contest version. Coding is not a forte. Creative writing, math, scientific for a decent average level is okay. Let me know your thoughts in the comments again. If you want to help out the channel, please become a member. Thank you for all the support.