Big Wins for Open Source | TONs of New AI Projects! (All Open)

Open source AI vs closed source AI. What do you think's better? This has been a big debate in the AI space since I can remember, to be honest with you. And well, the longer I sit back and watch AI develop, the more I feel like this intellectual debate about which is better doesn't really matter because it seems like open source AI is just going to win. It's been a long road. Back in the day, closed source AI was king. I mean the models were so much better. They're way larger and the smaller open source projects were fun but didn't really compete. But lo and behold, stable diffusion, llama, et cetera. And oftentimes open source models these days are doing things that closed source models still can't do. Regardless, in all of these AI sub niches the gap is closing quick. And especially in the last few days there have been some large open source drops that I think you should know about. So folks, let's see where open source AI is at first. I want to start with some rumors. Take these with a grain of salt as always. They're not rooted in truths, they're just some things that are circulating about the upcoming deep sea r2 model. I'm sure I don't need to explain this to you guys, But Deep Sea R1 was a big deal. They essentially replicated and reproduced OpenAI's cutting edge reasoning models in an astoundingly short amount of time and then of course released them open source so everyone had access. Now reasoning models are the norm for everybody. Anyways, DeepSeek R2 is a highly anticipated model because of R1's massive launch. Let's take a look at these rumors. 1.2 trillion parameter model which is absolutely massive. Only 78 billion of the parameters would be active at one time, and we could assume there'd be distills and smaller models as well because a 78 billion active parameter model, this thing would definitely take some compute to run. But somehow the claims here are also that it's 97% cheaper than GPT4O, which is already a cheap model. From OpenAI standpoint it's just 7 cents in and we're assuming per million tokens and 27 cents cents out per a million tokens. Again, chubby states. If this is true, we're seeing a new behemoth rising and I would agree with that. But again, these are just rumors. Deepseek R1 already kind of pulled off the impossible with how fast they replicated reasoning, but to scale it up to this size and we're assuming benchmarks that are increasing with that scale and pricing this cheap. If these rumors are true, the large language model gap between closed source and open source AI might be closing this year. And I'm sure a lot of people would say for the most general average user cases, it already has this one right here is just a quick recap. In early 2024, Sora, which wasn't even released publicly, this was a closed source model that OpenAI was teasing us early last year blew up. It could generate one minute videos and they were very high quality. Now, as Cocktail Peanut points out, in 2025 you can casually generate two minute long videos on your PC for free. And that's no exaggeration because Frame Pack does this made by the same dude who made Control Net generates videos progressively and you can do long videos with just six gigabytes of vram. And yeah, here's Cocktail Peanuts, a two minute long video. To be fair, it's not too bad. And as the video continues, sure things might get a little bit more mushy and strange with the motion. It's not perfect, but just crossing this barrier long generations on low VRAM is a feat in itself. Today's video is sponsored by none other than me. Hey, if you like the content I make here on the matvid Pro AI channel, you should definitely check out the Discord Server. General chat is buzzing 24. 7. People are constantly talking about the latest AI leaks, trying out the latest models, and of course making many AI images of yours truly. And if you want to share your AI creations, there really isn't a better place. AI images, video, songs, you name it, we got it. Whatever you make, someone's ready to hype it up or give you constructive tweaks. And harkening back to my roots, we're actually spitting up an official Mat vid Pro Minecraft server. This Saturday at 2pm EST, I'll be live streaming the Minecraft Server Build event where anyone from the Discord Server with the OG role can help build spawn. Regardless, I'm hanging out in the server almost every single day, testing stuff out before it hits the channel and oftentimes answering questions in real time. If that sounds fun links down below, jump right in and say hi. There's a bunch of other Discord Server features that I didn't even mention and we've got a fantastic tight knit community on there. Alright, well thanks to myself for sponsoring this video. Now back to your regularly scheduled content. Let's take a look at a voice model. This one was trending on hugging face with over 3000 stars on GitHub in just 24 hours DIA. This is a text to speech model that actually can listen to your instructions. And again, closed source models were doing this and it was like a big deal not that long ago at all. But this guy is fully open source and it does it well. Take a listen to this sucker.

We just hit over 3k stars on GitHub. Not to mention number two trending on hugging face happened in just under 24 hours. That's insane. Let's fucking go. Thanks for all the support. If you haven't already, try daya out on GitHub and hugging face. Excited to see what you create. This was gnarly labs.

We just hit over 3k very quick, but it sounds human for sure. Great emotion, great tone. And I assume these are cloned voices as well. And yeah, just like the last one we took a look at. Frame pack. This is fully Open source. Apache 2.0 doesn't get much more open than that. Sure we've had great text to speech from open source providers in the past, but the full control over the scripts and the voices, adding things in like laughs or an excited tone, that's something that closed source still kind of struggles with. Not only can you run this thing locally on a computer because it is open source, but you can also try it for free on Hugging Face. Just make an account log in and the link will be in the description. Let's do a nice funny prompt in brackets like that. S1 for speaker one. Kind of a weird script, but we need a weird one to test it properly. Properly. All right. 19 seconds.

My Lord.

Can you believe it?

I think that I finally lost my damn mind.

What do you mean? Your mind was my favorite one to peruse through late at night.

What.

Did not do the shivering.

How would you even go about do something like that? You know, I've got this whole place wiretapped.

A little bit of a screwy generation, but if you learn to prompt this thing well, it definitely could be useful. I'm gonna give it a quick rerun.

My lor. Can you believe it? I think that I finally lost my damn mind. What do you mean? Your mind was my favorite one to.

Peruse through late at night.

What? How would you even go about do something like that?

You know I've got this whole place wiretapped, right? This thing could be so random. They do have some quite literal fun examples though on this notion page. So here is theirs, the open source dia.

Oh fire. Oh my goodness. What's the Procedure. What do we do, people? The smoke could be coming through an air duct. Oh my God. Okay, it's happening. Everybody stay calm. What's the procedure? Everybody stay fucking calm. Everybody fucking calm down. No, no. If you touch the handle, if it's hot, there might be a fire down the hallway.

Perfect. I love it. It definitely sounds pretty real. There's something like memey or just fun about the way it screams and highlights emotion. Here's 11 labs. Traditional closed source text to speech. Oh, fire. Oh my goodness. What's the procedure? What do we do, people? The smoke could be coming through an air duct.

Oh my God. Okay, it's happening. Everybody stay calm.

What's the procedure?

Everybody stay fucking calm. Everybody fucking calm down.

No, no. If you touch the handle, if it's hot, there might be a fire down the hallway. So definitely less emotion. More robotic, but more clear than DIA, I would say. Still, I think I prefer DIA's a lot. Alright, we're gonna jump back into video generation. Dreaming Tulp points out that we're so close to realistic generated human motion in videos and shows this off as an example. Turns out this is open Source Realist Dance DIT. This builds upon the already open source WAN 2.1 model to generate more realistic human motion in videos. And if you've spent any time with literally any AI video generation model, you'll know that this is all quite impressive. Good human motion, realistic human movements. I mean that is crazy right there. They are not easy for these models to reproduce, but if you train them specifically and with this open source project, we're able to generate all kinds of unique human motion that looks straight up real in a lot of use cases. I mean, this one right here, this is pretty crazy. Like I don't even know how you would prompt that. If I saw that, I would just assume it's real. These are all AI generated though. So totally insane. Some slow motion video. This guy dancing. Like normally we'd see limbs and everything getting all mixed up and stuff, but this thing just nails realistic human motion. And again, would you look at that. Apache 2.0 licensed and like I said, Realist Dance here builds off of WAN 2.1 and WAN 2.1 was a great video generation model that again was already open source prior. That's the thing with these open source AI models, they actually build off of each other and help each other. One open source model can spawn another open source project that improves that model drastically in a certain specific use case. Here's how the finger dancing was done. You can See, it does require a mapped out humanoid figure of some form. We've seen these mapped out skeletal human figures and AI generation for a while now, but to do it this clean and this convincing with video generation is a different story entirely. And I have a feeling a big part of the reason it looks so good is down to the fact that it is built off of WAN 2.1 ultimate realism, controllability and human movement combined with all the benefits we traditionally get out of our favorite video generation models and open for anybody. And speaking of wan, they're constantly updating as well. This is a little old April 22, but the Open source WAN video generation model is free for all users on their website. Sure it's relaxed mode, so your generations might take a little bit longer, but you use no credits and you have unlimited AI video generation for free and it's pretty darn high quality. Check out some of these examples to see for yourself. While yes, you can also run this on a pretty beefy computer, a lot of people don't have computers and still want the benefits of open source models. Thankfully, WAN is just offering it for free and again, you're not going to find free stuff like this with a lot of closed source companies. They're almost always going to have credits or a very limited amount of free generations per month or per day. Unlimited relaxed free generations is a huge difference and I don't think that you would see this if WAN 2.1 wasn't open source. The fact that Alibaba released WAN open source in the first place means that the whole point of making the model wasn't to commercialize it on a website. While they still offer memberships for people who want to pay for faster generations, the model is open source, so having a relaxed free mode is just par for the course. Next one this is pretty interesting. It is an RL trained email research agent that searches your inbox and answers questions more accurately, faster and cheaper than OpenAI's O3 for example O3 of course being a proprietary closed source model from OpenAI. Now let's be honest, O3 also can do a lot more things that this RT model can't do and it's not designed specifically for this email task. But it is impressive that the open source RT here completes the task better than O3 while being way way lighter. Some really interesting notes in here as well, saying they were inspired by OpenAI deep research which showed how effective reinforcement learning can be to teach an agent a research task in particular. And for this one of course it's answering the questions about emails. The goal with RT and it seems like they accomplished this was to replicate similar performance to leading cutting edge AIs with open data and code. It's way more accurate, 96%, which is pretty darn good, five times less latency and 64 times cheaper. But I think a lot of people would probably say using O3 or even 04 mini is overkill for this particular use case. Anyways, they used 500,000 Enron emails and used GPT 4.1 to spin up some realistic question and answer pairs. Basically built their own synthetic data set. Not bad for a quick and direct model trained on a specific task, but yeah, full open source code and weights can be found on their blog all right, here's probably the largest piece of news you're going to find in today's video. Quin 3 is now available. It's already in LM Studio as you see here, so if you've got a GPU or a Mac, you can get LM Studio up and running and start downloading these models. You can see they've got a huge wide variety from this tiny little 0.6 billion parameter model up to the 235 billion MOE model. So plenty of choices depending on the hardware you're running and well, Quinn 3 is just pretty darn solid. They've caught up to deep seq with this one and it's better than as llama 4 saying it outperforms models like Grok 3 and Gemini 2.5 Pro. I don't know if I would claim that, but we do have some fun examples from God of prompt in this first up, complex reasoning. Quinn 3 solves a classic logic puzzle step by step without rushing. Each step was reasoned out clearly leading to the right answer with a full explanation, no hallucination or skipping steps. Keep in mind this is the fat 235B model. Pretty simple logic question. You can see the model starts thinking and it goes and analyzes the problem, breaks it down step by step. It does so pretty quickly, especially for the size of model we're looking at here. And it even goes ahead and double checks its work, lays out where the confusion is, gives you an answer and it's pretty well rounded, noting that when you mess up this logic problem, it's a interpretation of the standard English language phrasing. The fact that the language model was able to pick up on that is pretty impressive. It's also pretty diversely multilingual, which is great to see. Sometimes these open source models struggle with only supporting a few languages for the most part, but this one seems to be pretty impressive and well rounded. And of course guys, with these open source models you can just try them for free. A lot of times when you release a model open source you expect that a lot of people are just going to run it for free on their own GPUs at home. So well, oftentimes the demos to let you try them out are also free. I've given this thing a really difficult research prompt. It's been Basically a top 50 snack foods worldwide, ordered by total number of individual units eaten per calendar year. You can look at the rest of the prompt if you want all the details, but it quickly spins up a pretty darn decent result. Man Bread slices, grapes, bananas, apples, potato chips gives us a nice detailed table like we asked. To be fair, I don't think this is answering as good as even OpenAI's O3 model or Deep Research would. Here's a Deep Research's answer. 33 minutes of research to give us an answer. And also it is slightly different short answer. I think this research result is a little bit more accurate, but Deep Research really is the king when it comes to this stuff and Quen's model here really isn't too far behind.04 mini high or O3 Quen stop. I don't want to log in. There's so many benefits though with these open source LLMs. I mean the structures are getting really mature for running these locally, whether it's a Mac or a Windows P Bunny here runs the 235B MOE model at 28 tokens per second on his M2 Mac Ultra. And that's a huge model. Highly intelligent and it's doing great. Absolutely consumer hardware. Although this is pretty specialized because you have to have one of these special M series Macs regardless, For a single M2 Ultra Mac this is crazy. I would not be able to run this on my RTX 5090. Just don't have enough memory. And of course diving into the benchmarks for Quin model almost half the size of Llama 4 Maverick at 402 billion params, it outperforms Meta's Llama 4 on almost every single metric. As Chubby points out, it really doesn't look good for meta here. The llama 4 launch was interesting to say the least. And now Quinn 3 fully Apache 2.0 open source is kind of munching it and eating it for breakfast all in the same month. Wins across the board for MMLU GPQA GSM8K pick your benchmark here. And yeah, guys, it is important to reiterate here. These new Quin 3 models, unlike previous times, are under Apache 2.0. I mean, fully open source. You get your weights and your code. This release really is a pretty huge win for the open source community. And of course, it wouldn't be a large language model test without a little bit of rotating hexagon. It passes this test pretty darn easily. This isn't the more complicated one with extra shapes rotating around in the middle with the balls spawning in right in the center and making it kind of like a maze, but it still does a pretty good job. And this is only the 30B model as well. This is a smaller one. Interestingly Enough though, almost 12,000 tokens thinking about how to create this, which is a little bit crazy. Open source AI is closing the gap in pretty much every single niche, but especially large language models where code output is becoming a huge use case. General use is also a huge use case. Just every day, like I use LLMs almost every single day. The gap is closing fast. I'm really interested to see what the rest of this year brings. I will say a very important note, and if I go back to my previous tweet here, Felix points this out. We've sort of crossed a threshold with our AI large language models of easy progress, and we've kind of hit a soft wall. New progress, new impressive capabilities are going to come from implementing new methods such as tool use, which is what OpenAI is pushing more for now. Scaling alone will not fully improve AI to its very limits. We've got to build new architectures, new systems, and invent new ways to harness the power the large language models bring us. Oh, good thoughts, Felix, good thoughts. And for our final open source model here, we've got an instant character model. Just simply upload a photo of your subject and now you have a consistent character in the image generator. So instant character here. It is open source, but the license here, you'll see that it's only for academic, research and education purposes, so no commercial or production purposes. And that sucks. Hopefully they change that in the future or another variant of this comes out. It's still great because it's open source, so we can see what's going on under the hood. We've got a good paper project here and we can test it out and use it for ourselves, but not to make money. You can see how it works though. I mean, you, you upload any photo, any character imaginable, and now you've got a fully customized model Essentially, you can even do text in it. You can have this lady now punching a punching bag on the middle of a train. It's not perfect. And we're gonna go through some of my testing because I did put this one through paces, obviously. I just uploaded a picture of myself for like a thumbnail. I'm just smiling here. A lot of times when I do a character reference, especially the more instantaneous training methods, it'll over exaggerate specific parts of my face. Or it might keep my smile here. You can see I'm smiling in this photo and every picture it generates, it might have me smiling because that's all it knows. This one honestly surprised me in the fact that it doesn't really work that way. It's pretty diverse. It's pretty useful. Here's the first image I generated. This was like the stock one. It was like a character riding a motorbike in the snow, something like that. It's got my gray sweatshirt, which is cool to see that it kept that. The face and hair kind of match. But I think like the nose, mouth, eyes doesn't really look too much like me. The skin color is right, the hair is correct, but the face not so much. But then I used my own prompt, of course. Eating a cheeseburger close up, professionally lit, that absolutely looks like me. It is kind of hiding a little bit of my mouth. But the eyes, the eyebrows, the hair that sunken in look that I am known for, it's definitely there and I'm enjoying a cheeseburger. It's not super clear. This could do with some upscaling and some touching up, especially if I was going to put it into a video model like VO2. But it's a. It's a solid instant character reference. Here's another generation. Same cheeseburger prompt again. Really impressed by how it got the eyes there going in for a nice bite of the burger. Even my hands look pretty darn consistent with my actual hands. The hair is awesome. On point for sure. And again, you can see that gray sweatshirt now. I was interested. I saw the gray sweatshirt. I'm like, is it gonna let me make a different kind of clothing? You know, put me in a blue shirt with red pants. I couldn't see my pants. They weren't in the original reference. So it got the red pants down, no problem. But that sweatshirt, man, it definitely was a little bit of a hang up. It's a little bit blue in this one. It kind of blue ified my sweatshirt, but. But try this one again. You can see it did over exaggerate some of the facial features, like especially the bags under my eyes. I will point out that I am adjusting the settings as I'm going through these generations. So this one I think was default, but this one I had increased the scale up here and as you increase this scale parameter, the more it tries to make the end result look like the reference character and that can lead to over exaggerations like you see in this. Now if you go the other way around and lower this scale, you'll get images like this where the guy kind of has some similar attributes to me, but for the most part he looks nothing like me. This is a completely different person. I will say though, the quality of the image has improved though. Overall, I like the stylistic choices here, so that might be a trade off with this one. The more it looks like your actual character, the less clean your image looks. But sometimes people want to go for this more rugged style, especially if you're doing AI video. Also wanted to try something a little bit more fantastic. Here's me riding a giant tortoise. I think this one came out pretty good. It seems the farther away your character is, the less it will look like your reference character, but that's kind of to be expected. Overall, the anatomy is not too terrible and definitely still resembles me for the most part. If I was using this as an input image for AI video, I wouldn't be too disappointed. Here's another generation. This one was a little bit more screwed up. This was actually a really, really nice one. But this one rerun was not as good as open source personalization. I recommend using this one in conjunction with an AI video generator to use consistent characters in your stories. All for the price of nothing. Especially if you use it with wan. See, it really doesn't matter which way you look in the AI world, Open source is catching up to closed source, or at least competing with it at every single step of the way. As always, all links are going to be down in the description below. But yeah, this seems to be like the year where a lot of gaps are starting to be closed. The AI video kings might no longer be closed source. The LLM kings might no longer be closed source. We'll see how it all plays out. A lot of interesting thoughts in this video. I'd love to hear your feedback. Do you think that open source is the number one way to go? Or should it be a mixture of closed and open source models? What kind of future do you envision? Thanks so much for watching folks. I'll see you in the next video and goodbye.