# Meta's New AI Model Muse Spark Just Changed Everything (Fully Tested)

## Метаданные

- **Канал:** SkillCurb
- **YouTube:** https://www.youtube.com/watch?v=RVRwgKfV1zE
- **Дата:** 14.04.2026
- **Длительность:** 7:47
- **Просмотры:** 217

## Описание

In this video, I break down the brand-new Muse Spark LLM by Meta and explore what makes it one of the most exciting new models in the AI space. From real-world applications to hands-on testing, this walkthrough covers everything you need to know before trying it out yourself. If you're looking for a model that blends creativity, reasoning, and multimodal capabilities, this is definitely one to watch.

📦 Resources:
🔗 https://ai.meta.com/blog/introducing-muse-spark-msl/
🔗 https://www.meta.ai/
📽️ Video Timeline:
00:00 – Intro
00:44 – Applications
01:55 – How to Access
02:23 – MacOS Clone
03:33 – Sudoku
04:22 – Product Dashboard
05:31 – Frontend
05:55 – Visual Classification
07:02 – Final Thoughts

#AI #MetaAI #MuseSpark #LLM #ArtificialIntelligence #AItools #Developers #BuildInPublic #Tech #OpenSourceAI

## Содержание

### [0:00](https://www.youtube.com/watch?v=RVRwgKfV1zE) Intro

Meta is finally back in the large language model business and they have introduced their new model which is Muse Park, which is something that they built from scratch and they're trying to actually compete with state-of-the-art models like Opus 4. 6, Gemini 3. 1 Pro, and models like GPT 5. 4 and Grok 4. 2. And you can see in some of the benchmarks, it is completely beating all of these models. It is really good when it comes to multimodal reasoning and also quite decent when it comes to health and agentic tasks. I mean, the humanities law exam, it has scored 50. 2%, which is better than Gemini 3. 1 and GPT 5. 4. Although it still needs some more improvements to actually be called state-of-the-art model, but this is very good start. Now, here are some

### [0:44](https://www.youtube.com/watch?v=RVRwgKfV1zE&t=44s) Applications

of the applications. You can see it is very good when it comes to visual understanding and annotations. Like as you can see in the demo right here, it is annotating the objects in a refrigerator. Or in this application where you can see the muscle goods. As well as you can see, it built a really good Sudoku game using this image right over here. But still, it needs a lot of improvement. But in this video, we're going to deep dive into this and we're going to test this out. As well as we're going to talk about what are the future plans for this model and how you can access it. Now, as for the scaling axis, it is focused on three areas. Pre-training where they have made major improvements to increase the efficiency of this model. Then it comes to reinforcement learning where they have improved its accuracy as well as reliability through repetitive training and reinforcement. And then it comes to test-time reasoning where the model actually thinks before processing something or giving an output, which actually saves the cost and optimizes token use. We have some of the benchmarks right over here. How Amy 2025 evolution worked and also the humanities law exam. But that's it for the introduction and theory part. Now, let's talk about how you can access it. So

### [1:55](https://www.youtube.com/watch?v=RVRwgKfV1zE&t=115s) How to Access

for that, you have to go to meta. ai and you can access the model right here in the web chat. But as for the API, it is locked for now. I don't know why, but the model is developer locked and consumer ready. You can test it out over at meta. ai, but you cannot access the API and use it inside development environments right now. I will be letting you guys know as soon as they release the development version as well. But for now, let's dive right into the video and start testing this out. So

### [2:23](https://www.youtube.com/watch?v=RVRwgKfV1zE&t=143s) MacOS Clone

first off, I went on and asked it to create a clone for macOS and this is the browser-based macOS clone. The look is not exactly like a Mac operating system, but still, if you click on the apps, it actually works. So, you can see it's using the glass effect. If I can open up Safari right over here, we can open up chats as well. You can see it's functional. We can type messages here and then we have mails as well. Pretty good. We also have the image or photo gallery where we can open up any of these images. Let's say this one. Pretty good, right? And we can close these and also maximize. This is really good. We also have this notes app, the music player right over here. And this is the demo track it plays. The terminal, we also have calculator right over here. We have the settings tab and we also have a small version of spotlight or launchpad. So, we can search apps right over here. So, overall, this is a basic macOS clone, but I think this is really good when it comes to animation and designing. You can see that the window design is quite close to macOS and the apps and features are quite functional. The second app

### [3:33](https://www.youtube.com/watch?v=RVRwgKfV1zE&t=213s) Sudoku

that I developed using Meta's new model is this game. So, we're going to play Sudoku on the browser. So, this is the game. Let's start a new game and I can go ahead and play. So, if you guys don't know what Sudoku is, it's a game where you actually have to make sure that you have the complete series. So, we have 1 2 3 4 5, but 6 is missing. So, I'm just going to go ahead and say 6 and place that right here. That's it. That's how it works. So, once you fill out all the empty boxes, you're going to win the game. So, this is quite good. We have hints, we have notes, we have check right over here. We're just going to check and we have an undo. We got hints right over here and we can solve it if we want it. So, there you go. We solve and this is how it should look once you're done with the game. Third

### [4:22](https://www.youtube.com/watch?v=RVRwgKfV1zE&t=262s) Product Dashboard

experiment that I did with Muse Park was one of the best generations. So, I built this interactive dashboard for my product and it shows type search by default. So, we can do a lot here. It might look a bit overwhelming. Although I don't want this extra information right over here, I would better want a more cleaner view. So, I can change the finish to matte, glossy, and metallic. And it changes in real time pretty quickly. And if I hover over here, you can see what we have. So, this is the memory foam right over here. We have the active noise cancellation device. We have Bluetooth and we can change colors as well. This is really good. Also, we can change the studio lights. So, this is studio light. This is sunset. This is night and this is neon. So, we can pretty much go ahead and take a capture of the product right over here and post it over at social media as well. And we can also change the headset right over here and we can share as well. This is a really good dashboard, I would say. We can also export the analytics and we can give it to any AI to analyze the stats and let us know how we can improve the sales even further. Now, this was my third test with Muse Park. The next test

### [5:31](https://www.youtube.com/watch?v=RVRwgKfV1zE&t=331s) Frontend

I conducted was to check out its capability when it comes to UI designing and front end. So, we built a web page for a yoga house, which is called Prana House. You can see the glow effect and also the animation on this object. And if we scroll down, I really like the animations and this is the live studio. We can join the waitlist. So, overall, this is quite aesthetic, I would say.

### [5:55](https://www.youtube.com/watch?v=RVRwgKfV1zE&t=355s) Visual Classification

And in the last test, I wanted to check out its visual understanding and annotation capabilities as it was shown right over here in the applications. If you can see, so they were claiming that it's really good when it comes to visual annotations, visual classification, and identification. So, what we did is we went on and we built this interactive fridge explorer. So, I gave it an image and I asked it to make an interactive web page to identify and annotate the items present in the image. And you can see we have the interactive version right over here. Hover over the oranges, it gives me the name, a description, and also a freshness tip. And if I hover here, it gives me the name, also the description, and a freshness tip. We can also click on the items right over here. Like I want to go with orange juice. So, it's going to go and show this information box. If I say milk, you can see right there, green cabbage. Then we have tomatoes, we have chicken eggs, and mint. So, it's working completely fine. So, this was the last test that I really wanted to try out because they were talking a lot about how it's really good with visual annotations and classification. So, with that said, I

### [7:02](https://www.youtube.com/watch?v=RVRwgKfV1zE&t=422s) Final Thoughts

want to wrap this video up. I hope this video was helpful for you to understand how powerful Muse Park is and how they built model completely from scratch and they're trying to build it up to the standards of state-of-the-art models like Opus 4. 6, Gemini 3. 1 Pro, GPT 5. 4, and Grok 4. 2. And also, we talked about its scaling axis and how Meta is improving. If you found this video insightful, hit the like button, share your thoughts or experiences in the comments below, ring the notification bell to never miss out on our daily updates. Thanks for watching and we'll see you in the next video. Till then, stay curious and keep exploring.

---
*Источник: https://ekstraktznaniy.ru/video/50862*