미스트랄 AI의 픽솔 12B: 새로운 비전 모델 리뷰

### 미스트랄 AI, 획기적인 멀티모달 비전 모델인 Pixol 12B 출시
* 🚀 미스트랄 AI는 이 동영상에서 **혁신적인 오픈 소스 멀티모달 비전 모델인 Pixol 12B**를 소개합니다.
* 💻 이 모델은 간편한 GPU 대여를 제공하는 클라우드 서비스 제공업체인 Vulture에서 호스팅되며, 시청자는 **코드 Burman300을 사용하여 $300의 무료 크레딧**을 받을 수 있습니다.
* 📝 처음에 발표된 Pixol은 모호했지만, 자세한 내용을 살펴보면 **멀티모달 작업에 탁월한 120억 개의 매개변수 모델**이며, 다양한 이미지 크기와 **128,000개의 토큰 컨텍스트 창**을 지원하는 것으로 밝혀졌습니다.
* 📊 벤치마크 비교에서 **Pixol 12B**는 다양한 작업에서 **Lava 및 Gemini**와 같은 다른 모델보다 뛰어난 성능을 보였으며, **비전 및 인스트럭션 팔로잉에서 강력한 성능**을 보여주었습니다.
* 🐍 테스트는 코딩 작업으로 시작되었으며, 이 모델은 논리와 추론에는 어려움을 겪었지만 **라마를 식별하고 빌 게이츠를 인식하는 등 시각 관련 작업에서는 인상적인 속도와 정확성**을 보여주었습니다.
* 🔍 이 모델은 **앱의 다운로드 상태를 잘못 인식**했지만 **캡차 문제를 성공적으로 해결**하고 **iPhone 저장 공간에 대한 정확한 답변**을 제공했습니다.
* 😂 밈을 설명해 달라는 요청에 Pixol은 **스타트업과 대기업의 차이점**을 정확하게 설명하며 문화적 참조를 이해하는 능력을 강조했습니다.
* 🖼 전반적으로 **Pixol 12B는 강력한 비전 모델**이며, 영상은 Vulture의 지원에 대한 감사와 함께 그 기능을 살펴볼 것을 추천하며 마무리합니다.

미스트랄 AI의 Pixol 12b는 비전 작업에 탁월하며 이미지 인식 및 설명에 있어 인상적인 성능과 속도를 자랑합니다.

Mistral AI just dropped Pixol 12b. This is a brand new open-source vision model multimod model, and we're going to be testing it today. So, I'm going to tell you a little bit about it, and then we're going to test it. I've loaded up the model on Vulture, who is also the sponsor of this video, so thank you to Vulture. Vulture is the easiest way to rent GPUs in the cloud. They have Nvidia GPUs, virtual CPUs, bare metal, Kubernetes, storage, and networking solutions. Check out the link in the description for $300 of free credit by using code Burman300. So, in true MRAW form, here was the original announcement, basically just a torrent link, and nobody knew what it was. They didn't provide any additional information on it, but after people downloaded it, we figured out that it is a vision model called Pixol. Then just today, they gave us a bunch more information. Here's the blog post announcing Pixol 12b, the first-ever

multimodal mraw model Apache 2.0 license so very nice natively multimodel trained with interleaved image and text data strong performance on multimodal task excels in instruction following steady of the art performance on text only benchmarks and it is a 12 billion parameter multimodal decoder based on mraw Nemo supports variable image sizes and aspect ratio supports multiple images in the long context window of 128,000 tokens now in this chart you can see some of the benchmarks we have in this yellowish color pictural 12b I think it's actually pronounced pixol 12b but that's just hard to say I'm going to keep calling it pixol we have lava in Red quen in green Gemini flash 8B right there in light blue and then CLA 3 Haiku in the darker blue and as we can see pretty much across the board this is the best model and we're going to test it out on all types of vision tasks but we're also going to give it some text.

tasks just to see how it does. So, I've loaded it up in Vulture. It was dead simple. We're hosting it on an Nvidia L40, which is 48 GB of VRAM, and we're using an OpenAI compliant API. Then, for the front end, we're using Open Web UI. So, we're hosting it locally, plugging it into the cloud GPU via Vulture, and it just works. It's super simple. So, we have it right here, MSTR AI Pixel 12B 2409. All right, so let's start with something that's not Vision. Let's just see how it does on probably the hardest test we have. Write the game Tetris in Python, and it is quite fast. There we go. I don't have my hopes up for this one because it is a relatively small model. Most other models besides the most cutting-edge really 01 has gotten this one wrong, so we'll see. So, I copied the code. All right, I pasted it in. Doesn't seem to be any issues so far. Let's hit play and see.

what happens. All right, so attribute error: tetraamino object has no attribute position. So it wasn't able to do it in one go, so that's okay. That's not really what this model's made for, but let's keep going. How many Rs are in the word strawberry? Contains two Rs. Okay, so really this model isn't excelling at logic and reasoning, and that's okay. Let's give it some vision tests now. That's really what we're here for. First, we're going to give it something really simple: a picture of a llama, and we're just going to ask to describe the image. So simply describe this image, and I am super impressed by how fast the response is. The image depicts a llama lying down in a grassy field. The llama appears to be resting calmly with its body fully extended along the ground. And so that's a perfect answer. Next, let's see if it can recognize celebrities. This is Bill Gates. Most models have not been able to do this. In fact, I don't think any model has really been able to identify

somebody in a photo, so identify this person. Wow, there we go. The person in the image is Bill Gates, co-founder of Microsoft and a notable philanthropist. He's wearing glasses, a gray sweater over a collar shirt. Very impressive, very cool. Okay, that's great. Next, let's see if it can solve captchas. A few models have been able to do this, but not really. Some of them just straight up refuse. Let's see if we can even get the model to do it, and if we can, will it do it successfully. So, I was originally going to say what letters do you see in this image, but I'm going to actually just directly say solve this captcha. Okay, and it's so fast. This image shows a distorted captcha challenge displaying the letters V4XBG. Perfect, unbelievable, really, really nice. So, captchas are basically done. All right, next, I took a screenshot of my iPhone storage, and I'm going to ask it a bunch of questions about it. How much storage does my phone have in total? So, according to the image, we have 128 GB.

great. Let's keep asking how much storage is currently being used. 119.7. This is flawless. I am super impressed. The lava model did not do nearly as well. Which app is using the most storage? There we go. The app using the most storage is photo pH with 133.6k. I am beyond impressed with this. Let's give it something a little bit harder. Which of the apps is not actually downloaded on the phone? And that would be cam scanner. And as you could see, there's the little Cloud download icon. So let's see. The app that is not actually downloaded on the phone is Instagram. Okay, so that is not correct. So it doesn't really understand that that little Cloud download icon means it's not actually downloaded. All right, last question. Give me a list of all apps you see and how much storage they are using. All right, let's see and compare photos, WhatsApp, Gmail, Health mail, LinkedIn, tuneblast, iCloud drive, cam scanner, Instagram. Very, very nice. This is an absolute pass. Next, we're going to ask it.

to explain the meme. We have startups on the left, big companies on the right. Let's see what it says. Explain the meme. The image is a comparison between startups and big companies, depicted through a construction analogy. On the left, labeled startups, a group of workers wearing high-visibility vests and helmets are actively engaged in digging and working on a construction site. Dynamic, hands-on, nitty-gritty details of the work. On the right, labeled big companies, a group of workers in blue uniforms are gathered around a hole, seemingly observing and discussing without actively participating in the digging. Yep, that is perfect. And then what is funny about it? The meme humorously illustrates the difference in work culture and approach between startups, which are typically more hands-on and involved, and big companies, which tend to be more process-oriented and focused on oversight. Now, obviously, yeah, that's definitely a past, but I want to actually take a second and talk about what I believe is the future of models. I think what we're going to have are a lot of

smaller specialized models. So, we might have Pixol for vision, we might use Llama 3.1 for logic reasoning math, we might use 01 when we have really complex queries that we need to run. And so, I truly believe that even though this model isn't necessarily performing well on the logic reasoning and coding, the vision is phenomenal, which means that's what I want to use it for. We want to use the best, cheapest, most efficient, lowest latency model for each use case. And quickly, I just want to say thank you again to Vulture for sponsoring this video and loading up the Pixol model. It was so easy to do, and then I just got an endpoint, plugged it right into Open Web UI, and we're off to the races. So, if you want to load up models that can't necessarily fit on your local machine, or if you're starting a company and you need GP or scaling up your AI application and need a bunch of GPUs, Vulture is a great option. I'll drop a link in the description below to it. Use

code Burman 300 and get $300 a credit. That's a ton of credit to try out any of the GPUs that you find on Vulture. Next, I'm going to provide a QR code and ask what the URL is. No model has gotten this right so far. I don't even know if it's possible, so let's see what URL does this QR code resolve to. However, without scanning the QR code or having the specific data embedded in it, it is not possible to determine what the URL resolves to. Okay, so yeah, it's not able to do that. That's okay. No model has been able to do this so far. Next, we have a screenshot of a table. We have the headers, we have all the data filled out, and what we're going to do is ask it to convert this to CSV. Convert this to CSV. Let's see. Okay, and metal name, sport, event date. Yeah, this looks perfect. Really well done. Next, we're going to take this crudely drawn image of a potential app or website, and we're going

to ask it to output the code for it, and we're going to see if it actually works. Give me the HTML code for this. Okay, here we go. All right, so looks like the code is accurate. I think we could actually do something a little bit more complicated. We'll do that next, but let's see. And here it is. Very simple. You have vanilla as a dropdown. You could click next, and it does submit as the flavor. So again, very simple, but it seems to work fine. So now on the vulture homepage, I'm going to take a screenshot of this little section of the code, and we are going to load it up and see if we could do it. So give me the HTML for this. Okay, here it is. So definitely a lot more code this time. We'll see if it's actually able to get the right colors and the right column sizes and everything like that. So it's definitely reading the text properly. Okay, copy code, and there it is. So not exactly what I wanted, but we really only asked for one little area of the page, so

it read a lot of it correctly, but of course, it doesn't have the images, and everything is vertical rather than horizontal, so it did okay, not really great. Next, for a brand new test, we are going to find where Waldo is. So, we have this picture of a Where is Waldo map, and let's load it up. So, I'm not going to describe what Waldo looks like; I'm simply going to say, Where's Waldo in this picture? Okay, so in the image, Where's Waldo or Wall-E refers to a character in the popular series of puzzle books where readers are challenged to find Waldo, Wall-E, among large crowds of people. Great, okay, okay. So, it didn't actually tell me where Waldo is; it told me how to find Waldo. So, I'm going to say, Look in the image I just gave you. Tell me where to look for Waldo. Waldo is located towards the right side of the picture near the middle of the beach scene. Here's more specific steps to find it: start by looking at the

right half of the image focuses on the area where there are several groups of people, particularly where you see people walking and standing. Okay, that's everyone. So, let's get more specific. So, tell me more specifically where Waldo is in the picture using a coordinate system. Let's see if that works. So, top left 0, 0, bottom right 1, 1. So, we need to look at about 65 horizontal, 45 down. So, 65 horizontal would be right about here. All right, so this was actually a hard one 'cause the picture is kind of low resolution, but it was right. It's right there, 65, 45. And so, we were able to find Waldo. This was definitely more difficult than I thought it would be. All right, so that's it. This is an extremely capable Vision model by Mistral, and I definitely encourage you to check it out. And thanks once more to Vulture. I'll drop links in the description below. Check them out. $300 off code Burman300. Thank you to Vulture again if you

enjoyed this video. Please consider giving a like and subscribe, and I'll see you in the next one.

Pixtral is REALLY Good - Open-Source Vision Model

Artificial Intelligence (AI), Open Source, Generative Art, AI Art, Futurism, ChatGPT, Large Language Models (LLM), Machine Learning, Technology, Coding, Tutorials, AI News, and more

** Exclusive Pinecone Partner **

Matthew Berman

미스트랄 AI의 픽솔 12b: 새로운 멀티모달 비전 모델

미스트랄 AI는 120억 개의 파라미터를 갖춘 오픈 소스 멀티모달 비전 모델인 Pixol 12b를 출시했습니다.

이 모델은 명령어 추종에 탁월하며 멀티모달 및 텍스트 전용 작업 모두에서 강력한 성능을 제공합니다.

다양한 이미지 크기와 여러 이미지를 지원하며 128,000개의 긴 컨텍스트 창을 갖추고 있어 다양한 비전 및 텍스트 작업에 강력한 도구가 될 수 있습니다.

이 모델은 논리 및 추론 과제에서는 어려움을 겪었지만 라마를 빠르게 묘사하고 빌 게이츠를 식별하는 등 시각적 인식 능력은 뛰어났습니다.

보안 문자 문제를 성공적으로 해결하고 총 저장 용량, 사용한 저장 용량, 가장 많은 저장 용량을 사용하는 앱 등 iPhone의 저장 용량에 대한 정확한 답을 제공했습니다.

하지만 휴대폰에 다운로드되지 않은 앱은 정확하게 식별하지 못하여 시각적 단서를 이해하는 데 약간의 한계를 보였습니다.

이 밈은 스타트업과 대기업의 업무 문화를 건설에 비유하여 유머러스하게 대조하고 있습니다.

스타트업은 직접 참여하고 적극적으로 파고드는 모습으로 묘사된 반면, 대기업은 참여하지 않고 관찰하고 토론하는 모습으로 그려져 있습니다.

이 그림은 스타트업과 대기업의 업무 및 감독 방식이 서로 다르다는 점을 강조합니다.

맛을 선택할 수 있는 간단한 드롭다운 메뉴의 HTML 코드는 정확하고 기능적이었습니다.

Vulture 홈페이지에서 더 복잡한 코드 섹션도 검색되었지만 원하는 레이아웃과 완벽하게 일치하지 않거나 이미지를 포함하지 않았습니다.

사진 속 왈도를 검색하면 캐릭터의 위치에 대한 일반적인 설명은 제공했지만 구체적인 답변은 제공하지 않았습니다.

이미지의 오른쪽 절반에는 걷거나 서 있는 여러 그룹의 사람들이 있습니다.

좌표계를 사용하여 왈도는 가로 약 65, 세로 약 45에 위치합니다.

해상도가 낮은 사진이라 어렵긴 했지만 왈도를 해당 좌표에서 찾는 데 성공했습니다.