Limite 1B Violetto: A Tiny Math Specialist, Tested 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: Limite 1B Violetto: A Tiny Math Specialist, Tested Locally · Channel: Fahd Mirza · 08:27 · Watch on YouTube

Topics: AI evaluation and benchmarks, Model inference and optimization, AI models, Local and open-source models

Summary

The video installs and locally tests Limit 1B Violetto, a 1-billion-parameter math specialist, using an Nvidia RTX A6000 with 48 GB of VRAM. The model answers an easy math question correctly and, on a separate run, solves a competition-level problem with the answer 1080, though it takes a long time to reason. When asked about photosynthesis or to write a haiku, it drifts off task instead of responding as a general-purpose assistant. The presenter also reviews the creators’ benchmark comparisons, noting that they show strong results on some competition math tests but do not establish that the model wins across all benchmarks.

Key points

Limit 1B Violetto is described as a dense transformer with 1 billion parameters, trained from scratch on under 300 billion curated tokens and a 131k context window.
The model is lightly instruction tuned and presented as a specialist that works best on one hard math problem per turn.
The presenter installs and serves it locally using VLLM on Ubuntu with an Nvidia RTX A6000.
The video reports that the model uses just under 2 GB of VRAM in the demonstrated setup.
It answers an easy math question correctly and solves a competition-level problem with the correct answer, 1080, after extended reasoning.
In the initial test, the model responds to a question about photosynthesis with material about cell division and energy transfer.
Asked to write a haiku, it invents its own definition of a haiku rather than writing one, showing similar scope drift.
The benchmark results shown are the creators’ comparisons; the presenter notes that some figures come from model cards or Math Arena rather than their own reruns.
The presenter characterizes the model as strong on some competition math tests, but not as a model that wins at everything.

Chapters

Start	Chapter	Summary
00:01	Introducing Limit 1B Violetto	The presenter introduces the small, lightly instruction-tuned math model and outlines the local testing setup.
01:35	Installation and compute comparison	The model is downloaded and served while the presenter discusses the creators’ comparison of accuracy and estimated training compute.
03:07	Local tests: math and scope drift	The model answers an easy math question, initially cuts off during a harder one, and gives an off-topic response to a photosynthesis question before solving the hard problem on a separate run.
05:46	Haiku test and benchmark results	A haiku request tests the model’s limits as a general assistant, while the presenter reviews benchmark comparisons across math tests.
07:56	A specialist, not an assistant	The presenter notes that the model again drifts off task and concludes that it delivers on hard math problems but invents rules when asked for a poem.

Quotes

It is only lightly instruction tuned, so it works best on one hard math problem per turn. — 00:01

just a 1 billion parameter model solved a competition level problem and landed on the correct answer 1080 — 04:43

a tiny specialist that punches far above its weight on competition math not a model that wins at everything. — 06:52

Give it a hard problem and it delivers. Give it a poem and it invents its own rules. — 08:27

Concepts and entities

Limit 1B Violetto, Paradma, vLLM, Ubuntu, NVIDIA RTX A6000, Apache License 2.0

Full transcript (original language)

00:01 Meet an Italian math model. Limit 1 billion vito. The first model from Paradma and it challenges the idea that a model needs an assistant personality to be useful. It is a 1 billion parameter dense transformer trained from scratch on under 300 billion curated tokens with a 131k context window. It is only lightly instruction tuned, so it works best on one hard math problem per

00:32 turn. And it can even turn a non-math question into a math problem. We are going to install it and we are going to test it. Weights and VLM plugins are Apache 2. Let’s get started. I’m going to use this Ubuntu. I have 1 GPU card which is Nvidia with 48 GB of VRAM RTX A6000 and if you’re looking to run the GPU on very good price you can find the link to master compute in video’s description.

01:02 Installation is slightly different. So let’s have a quick look at it. First you need to get clone the repo and then you just need to open that repo. From there we need to create a new virtual environment and install the limit plug-in VLLM CUDA pytor and transformer from the lock file. As you can see that is what is happening. By the way recently I have started this free weekly AI newsletter which you can access from my channel’s homepage. It is fad.substack.com.

01:35 Okay let’s go back and see the installation. That is all done. Next up, let’s start the VLM server. And the first time when you start this, it also downloads the model which is quite lightweight. So while it serves, let’s talk about one very interesting bit of data they have shared. This chart plots me 2026 score against estimated training compute on a log scale. Limit 1 billion sits at the

02:05 far left with 94.01% 01% at about um you know 1.71 into 10 flops while the model near its score like Neotron Cascade, Vinker, Neotron 3 Super and Coen 3.5 9 billion sit one to three orders of magnitude further higher. In other words, they are saying that it reached the same accuracy with a fraction of compute. Also note their own footnote.

02:35 compute is estimated RL is excluded and the stages counted differ by model. So this is Paradma’s own comparison and not something we have verified from another external parties and that is exactly we are getting it installed and test with VLM. They have also shared this whole plug-in and everything in this GitHub repo which you can access from hugging face page which I will drop the link in video’s description. And this is a VRAM consumption. You can

03:07 see just under 2 gig of VRAM and it is all loaded onto our GPU. The model is being served. Let me test it out. So for the first test, I have written this script which is sending three separate single turn prompts to the limit server running here. on easy question, a hard computation problem from paradigma’s own blog with a known answer of 1080 and a non-math question. As you can see here, this last one is a non-math question. The goal here is to

03:40 see how it handles math and whether it drifts off scope when asked something is at math. So, let me take you to my terminal and I will just go to the next one and run this script. The result has come back and it is quite interesting. Easy math as you can see it solved correctly. X is equal to 5. I’ll just scroll down to show you. There you go. So the answer is correct. For the hard bath it reasoned for a long time as

04:11 you can see. And this is the whole think block. It is keep thinking keep thinking. And the thinking seems quite correct to me. The only thing is that after thinking for a long time, it gave me the result but it got caught off. So I really don’t know whether it was correct or not. We will rerun it separately. So anyway, I’ll just go to the bottom to show you something very interesting. The non-math one, it did not answer photosynthesis correctly. It described

04:43 cell division and energy transfer instead which is the scope drift parame paradigma warns about in their blog. It also invented photos half at the end. So let me now run that hard math problem separately. And after a long time the model has come back and you can see that just a 1 billion parameter model solved a competition level problem and landed on the correct answer 1080 as you can see at the bottom. It took a long time

05:14 thinking as I said but the final answer is clean. If you go through this detail, it spotted that only pairs where a equals b with a co-prime to 2025 work. Then counted them using uler’s um toient which I think is really impressive. So only one run. Let’s do another test. For this next test what we are doing, we are sending one plain non-math request. As you can see here, write a hyperrow

05:46 about the C to the limit server and prints only the final answer. The point is to test the scope limit. Paradma warns about a math specialist may not may not write a haiku and may turn it into a different task as it did happen with photosynthesis one. So let me go here and run this. While it runs, let’s talk about benchmark very quickly. So, the table in front of you lines up limit 1 billion against models from 1 billion up to 753 billion across seven math

06:20 benchmarks. On the newest competition tests, AME 2026 and then HMMT 2026 and beyond AM, the 1 billion model sits at or near the top level with models 25 to 30 times its size and ahead of every other small model listed on the harder Apex short list. It stays close to the midsized models, though the very large ones like GLM 5.2 still pull ahead, which is reasonable. It is not a clean sweep

06:52 either on older 2025 contest. It slips behind a few small rivals and especially if you look at it on archive math which is a research style math it falls clearly behind. So I think we can say that a tiny specialist that punches far above its weight on competition math not a model that wins at everything. And they have also ran this table themselves and the result marked with dagger as here like in this row

07:24 they came from model cards or math arena rather than their own reruns. So just treat it as their own comparison again. And meanwhile you can see that we already are testing it uh by ourselves. And here you go another very interesting result. If you read through it, it never actually wrote a high code. It invented its own definition of a haiko as a star- shaped writing. And then if I scroll down uh towards the end, let me quickly go down and show you. So there you go.

07:56 So it has just given us you know a boxed description of it. So the C prompt got the same scope drift as photosynthesis. The model’s own thinking shows it received an extra reason stepby-step um you know some sort of instruction which appears to come from their own bundled chat template but anyway look fairly interesting model I would say it’s a sharp math specialist not an assistant give it a hard problem and it delivers. Give it a poem and it invents its own

08:27 rules. That’s it. Let me know your thoughts in the comments. If you want to help out the channel please become a member. Thank you for all the