🦃 ThursdAI - Thanksgiving special 25’ - Claude 4.5, Flux 2 & Z-image vs 🍌, MCP gets Apps + New DeepSeek!?

everybody to ThursdAI, Thanksgiving special.

November

Volkov. I'm on AI Evangelist with Weights, & Biases I have. Joining me here, Wolfram Raven Wolf, AI evangelist, and AIvaluator. he'll tell us how to pronounce this and Yam Peleg. Oh. and we are here to celebrate and give thanks for the bunch of AI releases this week, and not only this week, even from the past year. And as said in the beginning, only today I realize that every. Every Thanksgiving is a Thursday. And so this is our third Thanksgiving special with you all in a row. So we would like to invite you, those of you who tune into the live show to participate with us and show us happy Thanksgiving and tell us what kind of mashed potato gravy, whatever you like. we're in the Y show, but still it is a holiday. With that, I wouldn't wanna say, hi Wolfram. Hi Yam Wolfram. How are you doing?

So yeah, it's been another of those weeks, where you get so many

releases, you don't know where to start, what to test, and it's exciting. Feels the acceleration

a hundred percent.

And we will definitely do a quick cover. and yum. What about you, dude? What's new

man?

Opus is the thing. Opus is crazy. Opus is the thing. It's crazy. I got, I don't know if we have enough time for what I wanna say about Opus. I got so much, but yeah, Opus is a thing.

Opus is absolutely crazy.

We're gonna get to talk about Opus for sure. Wolfram, what is yours?

Oh, I will pick something.

Open waits open source. Yeah. Z-image turbo That is Apache license. So yeah, that is my highlight of the week, but also have to mention Nous and Video 1.5. if I have to pick one of those, the image is the more open ones or in the spirit of Thanksgiving. Thanks a lot for that. And I think we'll see a lot of it.

We have folks in the comments surveying Z image from Tongyi lab.

folks, Tongyi Lab from Alibaba is like everywhere and they already have Qwen image and Qwen image edit. And now there's a completely new kind of thing that comes and says, ah, you know what? We have a different one. so ZZ image Fast was also released. we'll definitely show here on the stage because it's bonkers. Honestly. It is bonkers. And I really wanna, show you guys. my highlight this week has been the highlight of last week. It's Nano Banana Pro. I, folks, I cannot stop playing Banana, banana Pro. And I think that you guys will see why here on the show. And if those of you who saw the thumbnail on YouTube, et cetera, you'll see it is just, I'm just sitting and in awe in discovering how much this model can do and how much I can throw at it. Pages and pages of prompts, just precise things. I built it into our automation. I'll show you, I'll show you the automation as well, where. you guys are familiar with my research agent, that does the Twitter research, et cetera. Now, Nana Bon Pro generates an infographic for every little release that we ha it's just mind blowing. Insane. we wanna also say hello and happy Thanksgiving to Nisten who does not celebrate Thanksgiving today. 'cause Canada celebrates them like a month before. So you already have Yeah,

no, I'm down for a second round of, ah, yeah, let's go second.

Alright. Alright, let's go. Yeah.

Stan, welcome to the show.

we're just discussing the top things all on our plates this week. The top dishes for this Thanksgiving from this week. Yeah. who cooked? I'm trying

out the, the prime intellect model on their site now if,

yeah. Just wanted to say, shout out to Prime Intellect. Intellect three. Yes. Very powerful model. Open source. I

actually

used the base that they made for that one, and, I found it very good. I found it to actually be the first, like minimal, amount of active parameters for a good agent that can work also with the cloud code harness. Yeah, there, there's, there's quite a bit there, but

yeah, so we will definitely say thanks to the

folks from prime intellect for releasing this very late last night. and thus not making it into my, beautiful infographic, TLDR they wanna show you, but it's there and they have their own infographic.

I just wanna say, shout out to the whale deep

seek Math V two was released. Ooh, this, we should cover it.

Yeah.

Yep.

I think there's a very good review from Alpen, from at Alpen Dale, on that because he tried to do the proof for it and, but yeah, we can cover that as it,

alright folks, I think it's time to get started.

If you are watching this live on Twitter, on YouTube, on LinkedIn, hey, one, two, nobody should be on LinkedIn on Thanksgiving. If you're watching us on LinkedIn, go on YouTube, please. why are you on LinkedIn? but if you are watching this, no matter where you are watching this from, thank you for joining us. We're the ThursdAI crew. when we have, been here every Thanksgiving to for the past what, three years. we covered, every major AI release, and talked with every major lab besides Tropic and Tropic release this week. Hand topic. We would love to have you on the show. if you're releasing some stuff, you are more than welcome to come and hype up about it. And we usually start the show with A-T-L-D-R. Recently I heard some feedback. I did the AI Thinkers in Denver, shout out to the Denver crew. And, recently I heard some feedback that some folks skip the TLDR because they like to listen to our whole conversation while they do dishes or house chores. And and I otherwise heard feedback from some other folks that say, Hey, TLDR is always listen too. 'cause you guys talk too much. We're here to serve both these people. So let's start with the TLDR super quick. this time the TLDR is gonna be a bit different. so stay tuned. Let's go TLDR and we will get started.

This is the TLDR. This is the chapter of our show where we talk about everything that happened in the world of ai. Super, super, super quick, so that you'll be up to date if you don't have the two hours to sit with us. Specifically on Thanksgiving, if your Turkey is burning and you wanna run to it and you wanna just get to up to date, without extra stuff. So TLDR is for you. And this week's TLDR is special because we have this beautiful, nano banana pro generated, I should say, infographic. look at this thing. we'll start with. Start with open source. as always a great So folks, here's our headliners. It decided what headliners will be like. I think so. I don't know if, uh, if, uh, uh, open the eye decided via the prompt or, or the, you know, nano Banana Pro n. Banana Pro is power. The Gini three. So actually, like you don't necessarily need the opening. Nevermind folks. Anthropic launched Cloud Opus 4.5, its Premier Intelligence. finally is with us. Let me make sure that, yeah. This is not distract the premier intelligence, philanthropic. It's number one on coding. It's a state-of-the-art on coding agents. It tops the Swyx, be verified, benchmarks, crashes, terminal tasks, and add and effort now for cheaper, smarter workflows. So you can select effort, reasoning stuff for, ops and Opus is really for the longest time for many folks. Opus was the top intelligence and San is it's little broader and folks have been waiting for Sunnet, to to upgrade. It gets a record 80.9% on Swyx bench verified, which is absolutely insane. topping GPT 5.1 and Gemini three Pro at 76 and 77%, 44% on terminal bench, hard, which is, not state of the art, but really high as well. And it does. and Tropic also mentioned something about they have this like super hard engineering task that they give folks. And this basically beat every candidate that they have on the very hard benchmark. So we're gonna chat about Opus, for a while, and, Nisten will, chime in on his, I have used it for coding stuff, hopefully. the other headline that we have is a very interesting one from, from last week. And we'll have an interview with the two folks that have been on the show before. To tell us about this MCP, the Model context Protocol Standard. Tropic just celebrated a full year off. Can you believe it was just one year of MCP, that every major lab adopted, including Google, uh, OpenAI, obviously tropic Microsoft. Every major lab adopted MCP. MCP is a layer for APIs for, for models to, to talk to each other, and it lacks ui. And now there's a standardized effort to add UI to MCP apps as a standard. It's called MCP ui. And the two folks who came up with it, IDO and Liad have been on the show before. If you guys remember, with Git MCP. And since then they've created the whole spec and now it's adopted by tons of companies. So I would love to have them on because, first they're friends. Once friends succeed, we wanna highlight this on the show. But also because, I think it's a very important effort. OpenAI Apps in Chat GPT we're supposed to be like UI based, so now they're gonna collaborate. I really want to hear from them what's that about. But also MCP is celebrating one year. And with this one year Anthropic when, they released Opus 4.5. They also released a bunch of updates to MCP, like the. the tool, search tool and the programmatic evaluation, I would love for them to chat with us about those updates to MCP as well. let's go a little bit, into open source as Wolffer mentioned before, Hoan released, HY Hoan released to release it this week, right? HY video, the new update, but also they released an OCR model. It's a 1 billion parameter beast with beats huge V LMS on OCR bench and dock, parsing. It's really cool. And OCR has been like, really on fire lately. And then Microsoft of all folks released Fara seven B. It's an on the ice on-device computer use agent, it, 7 billion parameter, computer use or qa that, gets 73.5% of web wager Voyager with like very high beating, larger models, obviously in efficiency and navigates real websites from screenshots. And, it's really cool. We're getting computer use in small form factors. A lot of open shots. Hunyuan, also released a video ion 1.5. It's an open video state of the art. its 8.3 billion, diffusion transformer. Did video model, small five to ten second, video clips on consumer GPUs, with clever attention tricks for speed. And, we can show one or two, video examples, but that's not the cool stuff about the video. This week. This week, LTX Studio released, retake It's a new video editing tool. We've talked about LTX as like a bunch of models, Lucy and some other models. this one's kind of Photoshop for videos. you can take any video and then basically recast or change the ending for smaller shorts of time, like change motions. It's really cool. we have to play with this. and have you guys heard on this? There is a video arena that shows video generations from different video models to different folks and they judge based on. Who kind of thinks which one is the best. there's a new one there and VO three and 3.1. And so two are like the top ones there for a while. And now there's a new video model that nobody knows who this is from. It's called Whisper Thunder in parenthesis called David, and it is beating every other one for people liking. it's crazy when we don't know who this is. we are continuing in our kind of image and video. this has been a very multimodal week for open source as well. Flux two really. It's a 32 billion parameter multi ref image, monster generator. The reason why it's 32 billion parameter is because it's backed by Mistral, as it's also a thinker kind of image. it looks like everybody understood that the more intelligence you put into a diffusion model with an additional, LLM behind it, the better. it's a, rectified flow model, up to 10 reference images you can provide to it with native four megapixel outputs. not sure why this had megapixel not 4K, but this is four megapixel, very good for, realism and sharp text and UI rendering. and then also the same day. Togi lab from Alibaba is Z image turbo, depending on how do you pronounce your zz, or zz. it's a sub one second generator diffusion transformer. That honestly looks better than flux, And the Z image, turbo, is fully open source, 6 billion parameters or smaller than flu as well. And I think it's also backed by a smaller lm. I'm not sure. generates photo realistic images is just eight steps with strong bilingual text and, distillation tricks as well. So that's insane. And, I'll show you an example of Z Turbo, definitely because it's like super fast and I've revive something in this week's buzz. We have a new update, and Jeremy Howard was asking for it as well. if you have trained any model fine tunes, you can, you know that you can train a full weights update or you can train LORAs. LORAs stands for lo low rec adaptations, smaller files that makes you not have to move the whole weights around. Basically. Now our Inference service supports, LORA adapters in the inference provider. Provider. You just provide the inference, you upload them, and then we will combine the main model with your LORA. It's really cool. I have a visual demo to show you and everybody who is fine tuning models is welcome, to try it. We, we've opened it up as a preview for just a limited set of models, but it's really cool and I have a cool down as well. and it works for any OpenAI compatible API. we also had from this week, that's not on my picture. prime Intellect released Intellect three, which is their model, it's a, it's a Finetune on top of GLM Air, I think. it's a 106 billion parameter MOE. And it tops the benchmarks. it beats the GLM air, it does the benchmarks in open source. 69% lifeguard bench and 90% so on AI E 24 and 25. And, and then we also have, yeah, that image we talked about and, we Nisten. Last thing is dik. Yeah, deep think was the last

thing.

Yeah. Deep seek, math prover,

which I don't have a link for or the summary.

we will not have deep seek folks. I think it's time for us to start the show with, with open source. There's a bunch of open source. yeah, and also OpenAI launch, the new UI mode for advanced voice mode, where you can talk to it and it's gonna in chat mode as well. and that's pretty cool. And also a shopping experience, that's also pretty cool. One just in time for, tomorrow's gonna be Black Friday. alright. We don't have any discounts for Black Friday folks. Thursday subscription is free, If you subscribe to Thursday, I you'll get it for free. so definitely do hopefully not to annoying with the subscribe subscription prompts. Let's get started with open source. Let's do it.

Open Source ai. Let's get it started. All right folks, let's get it started. What should we talk about first? What is the most important open source release of this week? Would you guys say, based on all that we have?

Intellect three prime intellect.

Shout out. I think

what is so special about internet three?

Ya. you wanna talk to us? To it? I'll pull up the kind of the notes as well while this happens.

first thing first.

You got, an incredibly powerful model, open source. it's a large, expensive, open source, heavily, post trained, I think the most, heavily post trained models we had information to of how exactly was it done to, to say I, I'm not sure, but I think it's the first. I think we had, we never got this level of information about post-training till this release. It's a hundred b mixture of expert model, state of the art for its size on math code reasoning. and just, you just get the entire package. You get a paper that tells you exactly all the infrastructure, everything about mid training, how they did post training, how they structured the entire thing in order to just deliver this, incredible model. All open source, go to hugging face, download it, play with it. It's absolutely crazy. And, yeah, they've been on the radar for quite a while. Always committed to open source, with the, no, especially, pushing the verifiers, package. Yeah. starting this, starting this movement as well, which, is also major and yeah, I think a huge shoutout. People giving us the stuff for free. Absolutely. Yeah, a

hundred percent.

So primary intellect, we've covered the previous intellect, one intellect too as well. prime inte is, the third one, 106 billion parameters, MOE, with 21 billion parameter active, a Finetune of GLM 4.5 air, and it beats the GLM 4.5 air, which was a very good. Instruction based model, already, and it beats them at their own game, with 90% on, what is this? AI E 24 90%, almost 90% AME 2025. GPK Diamond is very high as well. it's a good coding model, and, humanities last exam, it gets very close to deep tick R one, with just what, one fifth of the parameters. So it's very impressive post train. said, they built a bunch of stuff, on their platform for prime internet, which they have, RL training or environments verifiers, and the environments hub, prime sandboxes. They really used all of the tooling that they offer on their platform for this. And this is one of the top, open source releases now that is coming from a Western lab, which is great to see as well. some competition. it is a Finetune of, not Western model, so we're still waiting for base models. Hello Llama. you guys remember we talked about llama, but, it's still very impressive. Go. By the way,

you guys can all try it.

Is that the chat? Prime Intellect, ai?

Yeah.

And if you log in, you can also add images.

it's very good at, just dumping in screenshots from the terminal and, and asking it to do stuff. also in other languages, like stuff that's typically not in training data sets. I saw though it was pretty good too, for stuff like if you want to do distributed computing or if you want to do anything in Ang, I was able to just, dump stuff in and, have it go through it, step by step. The one thing I have not tried this is to use clot code directly with it. That's what I feel like it's going to be very good at because this also accepts images. So your entire workflow for cloud code, now you can use a very well-trained small model that runs, if you're run it in four bit, you're only gonna need about, what, 70 to, 80 gigabytes of, of VAM or so. So it will also run on most, MacBooks as long as they have, like 96 gigs or, so yeah, this is now is probably like the best single GPU model that, you can run cloud code with that GPU is gonna be expensive, but it is. this becomes, if you were to spend under $10,000 as a company and you wanted to run cloud code privately, this is the best one you can use right now. so that's, but that's why it's pretty interesting in that regard.

A hundred percent.

the other news is DeepSeek has released a new meth model and here we have a tweet from Clem de, Clem says, as far as I know, there isn't any chatbot or API that gives you access to any IMO 2025 gold medalist model. Not only does this change today, but you get download the weights with Apache two license open source release of DeepSeek math V two on hug and face. Imagine owning the brain of one of the best mathematicians in the world for free to explore it for research Finetune, optimize and run it on your own hardware. And he posted a evals here for mean proof scores or on CNML level problems by category. So Deeps seek math V two beats GP five thinking high and Gemini 2.5 Pro. They didn't compare to Gemini three, which is interesting. at the time of the Gemini 3 was already available. deeps usually does a very thorough job at testing, their models. What else do we know about this one? yeah, folks are calling out in the comments that we get raw chain of thoughts, visibility into these models. Unlike the models, like Gemini and OpenAI and gr, we used to have a full visibility into Raku no longer. so this is deep seek math.

There, there are now some, some issues with it.

People tried, they dropped their, their paper in it and a very hard proof. And, strangely enough, the issue is the context window of 168,000, tokens is, too small for some proofs. And, The model kind of, this is not a model you're gonna use for an anything like general purpose because it can get very hung up on very small details of a proof. And, al been estimated that this would have worked if he could run it for two hours, but then it would probably need like a budget of like over a million tokens for it to actually complete the proof.

Oh, wow.

So that's an interesting limitation there.

I think that could probably be fixed with the tokenize and changing the, not changing the sampler for the output weight. But again, for most people, this is not something you really want to try and, dump your coating stuff in, right now. And there are limitations on the proof side as well. I think those issues can be solved, but, yeah, it's just interesting to you Keep that in mind

a hundred percent.

and so thanks Nisten for this. we're moving forward in open source. and the next thing that I do want to chat about is ion, from Tencent, a big Chinese company. Uh, ion releases there ion OCR. It's a 1 billion parameter state of the art end-to-end OCR model dominating the benchmarks. It has, OCR bench dominance. Let's see here. with 860 points, shadowing que three 72 billion parameter que three vl and, and other like rivals. I wanna, actually, I think I have in their announcement. I think there's an eval there as well. Let's take a look. this is very impressive. So on hug and face, downloads, as you can see, spike to 25,000 downloads. This model came out like two days ago or three days ago. This is really impressive. let's take a look at the evals. so Omni Docs bench, if beats. Pretty much every other model, G 2.5, pro dot OCR r that we've talked about here. paddle, OCR as well. OCR models has been on RISE lately, and this one seems also a competitor. deeps CR it beats that with 94. we have OCR bench and VQA in translation as well. So very impressive model. Again for 1 billion parameter with the Apache two license. I believe it's like license, other, licenses other, so not Apache two, they have their own community license, which,

isn't that the license?

That excludes the Europe. Europe,

the license agreement does not apply in the European Union, United

Kingdom in South Korea and is expressly limited to the territory as defined below. Yes, sir. Uh, WoFo woo woo. but it's still a great model, And all right, I wanna talk about Farra folks. Have you seen Farra from Microsoft? This also came out like fairly early on this week, far is a computer used agent, with Web Voyager at 73.5, which is able to, browse around click things. It's a Finetune, I love this. Microsoft Fine Tunes. Qwen 2.5 EL and Qwen 2.5. 7 billion parameter vision language model and it runs on device and under like copilot plus PCs that they've launched in Microsoft, under 500 millisecond latency in computer use. And you can run this on your computer.

I hope I'll get time to, to test it out myself.

It's an extremely interesting, locally running model, to, to control the computer. But, I didn't have time, so I have nothing smart experience wise to tell you about this at this point.

I will say, do we know the folks from, browser base and, they have

tested this model as well on their eval, and we're very excited about this as well.

but you could, maybe tell us how you want to test it.

What would you use as a harness? Would you use it with cloud code or how would you test it?

Oh, yeah.

Look, cloud code is, is just brilliant. The harness itself, Cloud code as a harness. It's definitely, without a doubt, much more powerful than the other CLI, CLI tools to date. how would I test it? I would simply just use it as, just controlling my computer. Things like open the browser, go to X, get my bookmarks. I bookmarked something interesting. please get it. The, I don't know, the GitHub repository from a good bookmarks I bookmark today that's on its own. Something that is not easy for models to do, even though it's an extremely easy task to do that reliably, where you can not be in front of the computer and when you come back, you see the gida repository with your bookmark. that's al already a test, but I don't know, just use it every day. I think that's the real task.

I think, without a tool there.

This remains in the territory of toys for enthusiasts, let's say but something like. Atlas or comet or other kind of like tool built in with computer use or web browsing things are absolutely in the realm of this is a superpower. right. We're not done with open source folks. uh, Ian Uhon, also released a video model also in open source. it's 1.5, version lightweight. Did based open video model pushing, sort, quality and efficiency. we can take a look at the release notes from them. this is a 0.5 release. so onion, state of the art, visual quality in motion. we can take a look at their video super quick, and see what's up with them. HY video, 8 billion parameter open source. the is probably one of the top open source videos that we have with 10 a DP. quality looks very nice, cinematic visual. again, when we have VO two and we have, and we have all these models looking at something like this and saying, oh, okay, yeah, we've seen this before, but this is open source. 8 billion parameters, 8 billion parameters that you can, run on your local hardware or run like, run slowly on your local hardware. But definitely you can run this on your local hardware. it's absolutely great HY video. the other thing that we should mention absolutely is that when the model is open source, it's way less, censored than the production models, and so on. Sorry, you can't like, generate anymore some of the stuff that you could before, for example. And, V three definitely limits you in what you can, what you can generate. these models are fine tuneable as well. Many people learn how to, Finetune these models with custom lawyers and do. Whatever, that, whatever, complete the blanks yourself. So the open source models are definitely celebrated around the world. And so shout out to Tencent for two releases this week. Two. Beautiful.

Unfortunately, I have to say, since you mentioned open

source, it has the same license, so it'll not be celebrated in the UK and European Union and South Korea,

unfor.

So I fully agree with you, ham. It's great to have your perspective on the matter, but I think the people who use this for fine tuning don't care. I think the license only prevents you from using this at your own projects. If you are building something, and if you're building something, then you should really consider, Hey, the cost of hosting this model on my own service, what is the cost of running like a V 3.1? you are definitely considering those, and I don't know if that wins to be very honest. the quality of what we get from the main production models, is so high that only if you're like, fine tuning and we already see like a few fine tunes for this model. but I think for, the fact that you could run this and Finetune this and generate some videos and they have no way of seeing whether or not those videos were generated, I think the license is not gonna prevent tons of folks of using this, especially not folks who are not like incorporated. and also this opens the door for stuff like this, like distillation models that are faster to generate the same video segment as well. confi UI workflows, et cetera. So this is definitely great release. but yeah, shout out to Tencent, like folks, you should let Europe enjoy the fruits of your labor for sure. All right, moving on. We're moving on, in open source as well. I don't think the LTX studio retake is open source, is it? No. So we're not gonna mention this. what else is open source? So just the image models that we'll cover. And then we're gonna go to big labs. So let's cover these super quick. also, black Forest Labs, have released F Flux two. Lemme open this. This guy, Flux Two from Black Forest Labs. This is a 32 billion parameter next gen image generation model. Unified 32 billion parameter architecture based on qua, sorry, Mistral based on Mistral. Both of them I think belong to the same accelerator in a 16 Z Friends. so I don't know what error their mute nerve transformer is, so probably this is a glitch in infographic as well. but they have rectified Flow Transformer. the new and supported breakthrough features in Flux. Two is the multi reference editing up to 10 images and styles. So here on you guys can see the editor and the editor is showing this highlight thing So basically in the prompt now, it's not just text, So this model is text to image and image to image. When you go to a ref image, you can see that like this reach editor, understands. I think it's a file feature, not a flux feature, but you can specify, lighting effects. But you can also specify images. So you can see at this image, change this thing at this other image include only the face, et cetera. So it definitely, definitely works. It's not super fast, but here we have, an editing of a vase. So this is like a V white vase. And, we just said change the colors of the vase in close the living room setting. Visualize the gradient flowing from this color to that color. Add artistic carving text with a big font that on this, that says there's the eye in the middle. And, this image out, this image model outputs. yeah, it's pretty cool. As you guys can see,

this amazing actually with the variables you can use,

because I've been doing a lot of image editing and you prompt it. You just write the image with the thing and put it there. Yeah. And, having more control, that is a big unlock. Very cool. Have to try it after the show.

Yes, sir. definitely have to try this.

the dev version is on hug and face, non-commercial license because they have a commercial offering as well live on file and replicate, and they claim to win in coherence and details and edit, versus competitors. I actually wanna pull up these, these release notes because we have to face this and Jacob wrote this in comments that the flux to launch was rough. It was rough because they will launch this model that is a non-op source model, even though they dropped the model waits just a few days after nano Banana Pro from Google launched and just changed the category entirely end to end. which is a bummer because if you are non-commercial, why would you use an open source model that's based on. Mis trial when you can use the huge, incredible model that's based on Gemini three that was trained by Google and has barely any restrictions. if you are after open source, they got squeezed from the other end as well. We're gonna mention Z Image just a little bit after this. So definitely a rough release, for Black Forest Labs. shout out on their hard work. Anyway, there, they're, they're considering themselves open core. So basically they, because they started stable diffusion, but they also need to make money, they decided basically we pair Frontier capability with open research and open innovation, releasing powerful inspectable and composable open weight models to the community alongside robust production ready endpoints for teams that need scale. this is their approach. And, folks, we will not have continued innovation in the open source without this, like open source teams also need to make money. So we wanna shout out, flux, BFL. so you can see some image results here. folks comments on flux on BFL or flux specifically on the releases of open source, open core, et cetera. Feel free to chime in here.

Yeah, it's really unlucky that the timing, if it had c come

two weeks ago, for instance, it would be celebrated all over the place. Yeah. but it doesn't change the model. It's still the same model. I tested it. I like that very much. And I think the combination with Nana Banana, you can create a lot of reference images and then make your own Finetune for this. So I think they are not just in competition, but they can, enhance each other.

Hundred percent.

so three models. Generally the state-of-the-Art Flux Pro, available under their BFL and their API Flux Flex is available at a control over our model parameter, so you can remove number of steps lower to for speed. And the flux dev is the 32 billion parameter open weight model on top of Mistrial 32 BI believe. they also have Flux Kline, which is coming soon. It's an open source, fully Apache two model, distillation of flux based models. so didn't come yet, but it is gonna come soon. And the Flux VAE is a variational auto encoder that other, labs use absolutely including the Z image. So Flex V could be the biggest deal from this release. This is like the auto encoder that basically Nisten, and correct me if I'm wrong, correct me if I'm wrong, this turns images into tokens and this does the best jobs at taking this. when the model takes other images as reference that like visual encoder, variational auto encoder is something that you need to represent your images in latent space. so definitely Flux two VA e is going to help everybody else. and if we're talking about image models and we said that flux, BFL was squeezed from both sides. One from, in the beginning before they released, they got undercut with incredible model in nano Banana Pro from Gemini. The other kind of undercut was Alibaba releasing Z Image Turbo, which is a 6 billion parameter single stream diffusion transformer, with sub-second eight step generation per image. And that's Apache two license as well. and it visually is better. Which is mind blowing. it like visually is better for, realistic images and it can be controlled. And, they didn't release the base one. They will be released it soon. And the edit one, so flux is also a image to image Model Z image is only text to image for now, but they will release the image to image soon. and it has also accurate text rendering. As you guys can see on this image here, all of these things are generated, including all those details. gimme one second. Zoom in here. so you can see very impressive, text generation capabilities in an open source model. Again, we cannot compare this to Nano Banana Pro because Napp App Pro has all the power of Google CPUs behind this. And MA three, this is a 6 billion parameter image. And look at these outputs. Just look at this beautiful 6 billion parameter. You can run this on your, on your computer like very easily. if Nisten is involved, you can probably run this on a toaster. definitely huge release. I think it's Apache two. Here's the performance on Image Arena. Let me load this, image arena. Okay? what the hell AI Arena from Alibaba. Did you guys see this? This is, I clicked a screenshot and this is like the first time that I even see this, but there's an. There's an arena competitor from Alibaba in addition. Wow. Alright, so the leaderboard on, their arena nano Banana Pro. number one, Gemini 2.5 flash image preview. This is, nano banana, not the pro one. And then we have C Dream four, which we talked to you about. And then Z Image is the fourth one. It is the only open source one, C Dream four, I don't believe it is open source. So this is the top open source model for images. but it didn't include flux here yet. and I built a demo and I love showing you demos because I really wanted to, mention that like the durable part of it, it only takes eight steps, only takes eight steps. so if you guys remember, this is neil.fun. This is Infinite Craft. This is a game like this, or basically you have some context. you have some nodes here. the basic, earth, wind, water, and fire and you drag them together and, like water and fire makes steam and steam. Plus steam means cloud and cloud plus wind means storm. And storm plus Earth means, I don't know, tornado, right? This is the game. It's really fun, as the game is really fun. on behind the scenes. The LLM kind of combines the concepts. I was playing this game two years ago and I was like, Hey, this game misses some, imagery. And so I built, SDSL max, I built a Chrome extension that basically prepares a prompt based on these concepts on here. And, With image generation models re lately, it's really fun to play this. Now, specifically, this is the Z image, the new release because it runs under subsecond you can see how quickly this would, this will work when I add a new concept. You can see a little bit of a loading, and then I use cereus here, for very quick prom generation. So this would be a pristine snow plane stretches across the rugged Earth Field Valley, where alone, blah, blah, blah, blah. I also updated this game to tell a story. So every time you a, like at the penguin, it will tell a story from the previous one with this penguin. So you can see the image, the, just the loading of the image takes a second. But like the generation for this image took less than one point half seconds. we have a crown generation based on this, these concepts and an image that is served from foul. It was super quick,

really cool, impressive demo, and at the same time, also a cool game.

It is a very cool game.

I love playing this. this is a Chrome extension, that I will probably post because I think it's now ready. The, you guys can play with again. The shout out is to, to Neil. Neil Agrawal is the person here with Neil do fund. and, you can see the generation of the Z image, like it generated the pirate. This pirate looks suspiciously similar to a cinema pirate that we know and love. And, we can add a ghost pirate there as well. All

pirates look like that,

I'm sure.

Yeah. Yeah. Very similar to it.

It's not suspiciously that's the real thing, but yeah,

respectfully, that's the real thing.

so the thing that I wanted to highlight here is, this model, while

not perfect, can take a bunch of concepts. And the cool thing about this model, first of all, they have a prompt extension module where if you just, most users just type, gimme a pineapple. And, in order to get like a good example out of these like image models, you need to like actually define all of the things. I decided to just use a prompt engineering. Put it on Cereus because the fastest one, I did use WB inference for this as well. but Cereus speed was absolutely crazy. And, the thing that allows this game allows me to do is to explore like all kinds of concepts and how this model specifically kinda plays with them. Nana Banana is not feasible with this game, right? So Nana Banana is absolutely crazy beautiful, but Nana Banana Pro specifically, it will take what three, I know, two, two minutes to generate an image. And this game, like the whole point is like iteration super quick. So you can add concepts, like put concepts on top of other concepts. Ghost Pirate means captain, captain, and you play around with these, I don't know what Pine is. the more you kinda play with this, just adds a visual layer to this game. And I absolutely wanted to show you. Let's see what Captain Morgan is about. Captain Morgan. Somewhere. I don't know where Captain Morgan is here. so this is my game, demo that I want to show, in addition to. Yeah, there we go. Captain Morgan with a compass. And, the drink is from the previous generation. we are now moving to the second part of the show. I wanna cover, this week's bus. Super quick, what happened in open source. And we stay with us because we still have to talk about the Opus 4.5 as folks are very highly excited about, coding examples. So we're gonna have, a little interview with, MCP folks and then we're gonna cover the Big Labs news as well. but meanwhile, I have to award from our sponsor. This was Buzz and this week's, this week is very exciting. This.

Hey folks, welcome to this week's Buzz. According on ThursdAI where I talk about, this, the news from this week from Weights, & Biases, the only sponsor for Thursday. I, Weights & Biases from CoreWeave. And this week I have something super exciting to you with a demo because this week we launched something that's called LORA Inference. just as a reminder, we have several things for you if you're training your own models and we also have an inference service. Inference service means that you can use Weights, & Biases, that's based on Coreweave CPUs with very cheap to run open source models such as Qwen and Kim K two and DeepSeek R one and, GPT O OSS 120 B models that you usually cannot run on your computer. we host them for you on core with GPUs, the only platinum tier cloud AI hyperscaler on the semi analysis. GPU tier. so if you wanna run your workloads on that same power, the powers up OpenAI, we have the option for you and very, for very cheap. So this is the inference, but the inference, usually what happens is we take open source models, we put them as is on our inference, and we give you an endpoint OpenAI endpoint, and you can use them. However, sometimes you need more. And for that we have two things for you. We already talked about serverless rl, a feature from the open pipe team that we Weights & Biases have recently acquired, where you can Finetune models with very few new additions and build basically Finetune for your own purposes, for new behaviors, for different things. But what happens after server SRL, what happens is you generate either a full weights of a new model or something called an a lower rack adaptation. a layer of weights on top of the main weights. That is basically you can attach it to any big model, the one that you trained it on, So basically a easier way to Finetune, and it's a very good one. And now. If you have trained a Low Rank Adapation, what we now offer you is the ability to run that LoRa on our infant service by just mentioning this. It's honestly very simple. Here's how it looks. I'll show you a quick demo here. So basically all you need to do is to upload your lower weights as a w and D artifact. It's very easy to do. and then afterwards, all you have to do is just mention that artifact URL as your model name instead of mentioning just the model name. So here's an example and I wanted to do a super quick, like a very funny example. I don't know if you guys familiar with mocking SpongeBob as a meme, but basically mocking SpongeBob is, when you wanna maximum in is capitalized is crazy. so I built, th this is a Finetune that I built on top of Quin, 2.5 14 billion instruct model the model that's available on our inference. And when you say, Hey, basically one of them will respond like the base model, hello, how can I assist you today? And my Finetune, which is a i generated a Finetune of these capitalizations will answer the, like this. So this is like a mocking capitalization that the internet knows about. And basically all that happens here is I serve the model weights, the regular model weights when 2.5. And here in the marketing SpongeBob, I serve a Finetune, LoRa. So a very small adaptation. This took less than 25 minutes to Finetune on top of open pipe. And now I basically have my own customized model on top of WB inference. So if you are interested in this, we have, a link to the, lemme copy this link here, a link to the instructions of how to do this here. at the bottom. Let me add this as a banner wandb.me/lora_nb. If you're interested of how, of a, how to on, how to do this. You can, check it out here on LORA, notebook, and let us know if this works for you. And we're gonna add more and more models as we go forward. So this has been, this week's buzz and, let's go folks. Let's move on to the next step of the show, as we have two folks, that are. Repeat guests of the show and have been with us before. Let me add them to the stage. Ido and Liad, welcome folks. let me reintroduce you super quick. You guys have been on the show before, as we covered Git MCP and, since then you continued working on this MCP standard. Yeah. And this week you came with a bunch of other news where I wanted to give you a stage on the show to cover because I think it's very important. very cool. so how about, yeah, maybe you start with the announcement, like what exactly happened, what released this week and though we're gonna chat with you about some details.

Yeah, so actually, this weekend it was super, super exciting.

MCP ui, which Ido started, in May. and we've been working on it since then. It was finally standardized into, one unified spec called MCP apps, which actually unifies not just MCP UI but also the Chat GPT apps SDK and that creates this base layer, base standard that every chat app, everyone who wants to build and wants to consume chat apps can use this spec. And it's like this build once run everywhere. promise. So that's super exciting. It took a lot of work. it's a joint effort of MCP UI, OpenAI and Anthropic, and obviously the rest of the community and, we're very excited about it. it really feels like the beginning of the new web.

I have a bunch of questions about this, but, as far as I know, and

last time we ta talked with you guys here on stage, MCP is a protocol for models to talk with some APIs, right? And there has been a lot of con conversation about this and would love to chat with you about like cold mode as well. But basically this is like a protocol layer between, models and APIs. How is UI related in there and why is UI needed? could you like give us a sense of why this protocol needs an extension into the UI world?

Yeah, definitely.

so the first part is that, when it was released back in May, we had two problems. One was that everything was text-based and text is terrible. So we needed something that is more, know, accustomed the way that people work. And the second was that when actually companies came to adopt MCP, they had a problem. They completely lost their identity, they lost their user experience, and it was a lot, a huge blocker for adoption for MCP. so once we have MCP and you have that, common layer, it's suddenly easy to come and say, if you wanna take the web to the next level and you want to embed real experiences into agents, why not just extend MCP? when I looked at that just added, looking at the existing primitives of MCP and trying to see how to extend it without breaking anything. So it's easy to adopt. and then it just took off and I think the amazing community just made it happen.

if I'm understanding this correctly, and I really

wanna to try to paint the picture for the audience as well. what you're saying is if a company exposes their own offering via MCP, they basically get lost because the LLM that. Consumes this consumes other things. Just like outputs as text doesn't maybe mention the company name or anything. Like something like this. So like the offering that you have is getting lost, whereas if you support like some sort of ui, let's say Spotify, right? You can use Spotify's MCP to get your playlist and then show 'em in text. But Spotify has a very specific brand identity and they usually need UI for that. Okay. Okay. I'm with you. And then, somewhere in, was that two months ago, OpenAI decided to announce their kind of like the app store will have embedded visual apps within the I GPT interface and they had some sort of this was built on top of MCP. so are they're competing with the standard that you guys are proposing or they're gonna support it? Could you talk about their, their offering?

So I'll just say that, MCP apps, which is the standard that was

just released in the weekend, it's actually unification of, MCP UI and J GPT Abcess decay, abscess Decay. I, they're very similar. They were very similar to begin with. we stopped working on MCP ui, so we'd released it in May. We've been working since then. And, they released that, yeah, like you said, in, in October J GPT apps. So I released it and they, we've been working with them to try to take the best of both words and create this, MCP app standard. So to avoid this, iOS versus Android or browser war, of definitions. Yeah.

how does this actually work?

So we basically, folks who are listened, they have a, some understanding of how MCP works. It just super briefly, there's MCP clients, that are the kind of deal to talk to the user. The MCP servers are the tools, the developers in different labs, database, Spotify, whatever. They expose their offering instead of an API, but very specific, Hey, model, here's how to use our APIs. Where, how does the UI layer, how does the app layer plug in there?

So it basically builds on top of that existing layer?

the idea there is that there are tools that expose ui. they are linked to resources that are actually the UI templates. Once, the model wants to display a certain application, it calls that tool, and then it can fetch that resource and just inject the data from the tool into that template. And then you can have, one static, kind of application template with different, data each time you call it. which makes it very dynamic and easy to manipulate.

And when you say template, what exactly do you mean?

Is that like a React framework type thing? Is that like web components? what do exactly do you mean by template?

So at the moment it's just a static HTML.

there were other methods that, that we explored in the beginning, like external URL, which is actually just taking an URL and rendering it inside an iframe. and we had remote down, which is a whole concept in itself, which is cool. but HTML is really like the easiest thing to do. like once everyone agreed on it, then it was easy to come and say, let's move forward. so it just taking, an iframe, double iframe say nested I frame, and just pushing the HTML into it.

So it's an HTML and responsive reactive as well.

So the LLM would know what the user did there, which buttons the user press, and what the actions the user took. Could you talk about the insight that DLM has within the, app?

Sure.

Yeah, that's a great question. so MCP UI is really, or MCP app, is really part of two things. There's one actually, how do you, take the UI and move it like the HTML. The second one is how does the UI communicate with the host? so the first action is, from the guest ui, into the host. You have a bunch of different action types. So the, if you think of it as a spectrum, you have, notify, which is the lowest one, which it says something happened within the ui, I want the agent to know about it. You can add it to the context. It can. Now, let's say, like I said, let's say the user clicked on something, but it, I just wanna know that it clicked it not more than that. and then you have tool calls. So if you want actually want to like, drive the agent flow forward and or actually communicate with your backend, you can just call that tool and the host will execute the MCP tool call on your behalf. or you can actually even just prompt, have a follow up, as if the user typed it, which really gives you like the full flow.

That's awesome.

LIAD, just stepping up from the super technical explanation of how this works, how do you guys get to the point where all major labs start to talk about this and support this? Could you gimme like a brief overview of like how this happened? you guys already previously explored with GI MCP and, and then take us from there from the last time that you visited the show through GI MCP, through MCP ui, and now MCP apps. Could you tell us, like, how do you get to a point where you guys collaborating with philanthropic, with OpenAI, on a global standard that could be like the next iteration of the web?

yeah.

Last time we talked about GT MCP, it feels ages ago in, MCP terms. and, I think back then we already understood that, the importance of, user interactions and getting feedback and displaying data for the user. ido, worked on, MCP UI and released, like this, UI over MCP, package li library around May. and, we've been working on it together since then in parallel with Git MCP, and it really caught on in with the community. early adopters obviously, really liked it because MCP was going up. People wanted this, to experiment with this UI over MCP kind of thing. But when you're building a standard, the first question that everyone asks you is who supports it? And it's like the chicken and the egg problem. You need servers to build MCP UI, and you need hosts to support MCP UI from the host side. so each of these sides ask, okay, who supports it on the other side? so we had this, it's not like it was a lot of hard work, but, I've been, I was working Shopify back then. And I managed to get Shopify to adopt MCP UI. So every Shopify, shop, every Shopify online store actually, returned MCP UI components. It was back in, July. they had this big agent kit release. and this really gave a push to the adoption of MCP UI we e We saw interest and adoption from Goose, from Postman, from 11 labs, from hugging face, from, very big companies. The, these were not the largest source, but still big companies. but the, and we were talking about it. We were, advocating it. We built a lot, a community. We had, lot of, Very, important influencers advocating for it and champion it. and then, we were in the MCP dev summit in London. We gave a talk about MCP UI, everyone was super excited and came to talk with us about it. And two days later, open released, the G GPT apps as decay, which is, which was very similar to c to MCP ui. Obviously, it builds on the same vision and concepts of, this, fragmented web and mini apps that are inside of one unified chat, chat interface. A lot of the technical details were the same. There were several, small differences. and, and then we started, they actually, they approached us a few hours afterwards, and they wanted to try to explore this kind of a unified spec in order not to be the iOS of the world. But, but again, Che GTI holds most of the consumer traffic. most of the companies were rushing to build Chat GPT apps, which was like an Im proprietary standard. so we try to think what we can do to make this like unified, generic spec. we were in good touch with Tropic already because of, of our action in the UI working group, in the MCP committee. so we've been working together and I think at some point it was clear, for Tropic as well, that we need as a community to do something we need to align on something, be before we'll have this one proprietary spec that, that will take over. and, and that's what happened in the past weeks and, maybe I'll let you though, continue because, yeah. I think, like the, yeah, I think the amazing thing here was like the realization that we can't get into that iOS versus Android space because everyone will lose, to have this kind of new world of apps. Like how do you actually distribute your software, and get the community to agree, is amazing. And I think it's a real testament to CPS community. like I think as soon as, Probably one or two weeks after MCP I was released. It was already a working group. so really kinda to see the whole community going together is amazing. And yeah, I think going forward after, philanthropic and opening AI and MCP UI have banded together to MCP apps. there's still work to do. We still need to finalize the s we still need to get the community's approval. but after that, I think there's so many stuff that we're looking into and so many new additions and, even collaborating with web MCP and going to a lot of different directions that are super exciting. yeah. So I think, it's up to the community to really come together and help us do it.

That's, folks, I wanted to highlight how important the work that

you are doing is, and how far along you guys both came when you just started like playing with some, oh, this is cool. And then from this you're like, shepherding a global standard between like huge, companies is very important because I think that, it's one thing for a company to build, like a Chat GPT app, it's completely another thing to have it then supported across everything. And I think this is what, where standards are coming from. And then Tropic basically mentioned this as well. And so this week, together with Opus 4.5, which we haven't talked about on the show yet, we'll talk in a second. Tropic also released some MCP additions, and I wanted to see maybe if I have two MCP, spec people here, if you guys are familiar with those and wanna talk about those a little bit, because I would love to hear expert opinion into how do people actually, start using these tools very smartly. would love to hear from you, if you are up to date with those updates. what kind of on topic suggests as a way to move forward in this spec?

yeah.

So I will say that we're not part of, any tool rated work groups. from our basic understanding, there are two big, additions. One is just having that pattern of a tool that searches for other tools as a progressive discovery of tools is one of the big things that philanthropic just, highlighted. that's a pattern that already existed, but now it's semi officially adopted. there's also asking tool calling, which is another thing, happening in the background. and I think, the third thing is really that adoption of code modes, or I think as it probably calls it, programmatic institution. Yeah. Something like that. Yeah. So I think. Those probably make a huge difference. I know even if we look at, one as how effective the models are, but, I think the other one is just how more, how much, it's super efficient to just run everything, without pushing everything into the context. I think it's interesting to see how that would work with advanced features like MCP apps. it's not entirely clear when do you actually display an app when you just run like a bunch of API calls, one after the other. so yeah, I think there's a lot still to, to uncover, but it's definitely exciting. I, I think to continue the lessons that is Ido said, I think something that we was very clear in the past two months is that all of those changes in MCP apps are meant to, build apps in, chat, chat environments. So not just presenting UI or, do, a tool called, but like full, functional apps end to end. when you have this like a code mode or code execution, it only handles the execution of tools. But we still need this, end-to-end solution for apps, And this, for this MCP really shines.

Yep.

So I, I think, just to to summarize the excited excitement development or we'll very soon see, LM chatbots stop being just like a textual kind of response. We already saw this with multimodality, where you can like talk to JGBT and get the response back to you. now they're bringing this like in, into the ui, but with apps, they definitely said, Hey, this is a third duration we're gonna try to take on Apple, basically. W with apps within our ecosystem, we have almost a billion users on a weekly basis on the ecosystem. Users need buttons to click and the companies need to be presented there. so this is what you guys are ushering, as well in the global standard. who adopts this at this point? Could you speak about like adoption and kind of demos and how, where now MCP apps are supported? who has an MCP app? where can people go and try out right now? what it looks like?

Yeah, so they can visit, mcpui.dev.

there's already a website, it was for MCP y, but there's already like a bunch of, supporting, clients. so some examples are Goose, which was the earliest adopter. and Goose is doing an amazing job of just, working with things. postman have an agent. they're, fast agent robot, like Tan Bot and a bunch of others. And obviously Chat GPT. and there are a bunch of other hosts, coming soon. yeah, do wanna take the servers. yeah. So obviously we have, Shopify. So it was, we mentioned supporting, supporting MCP UI and not going to support us. So MCP apps, like we said, hugging phase as well. we have, we see a lot of, companies focused around productivity and work, that are, looking into getting into, MCP apps and all the companies or the providers that are supporting apps, chat fee apps as decay, will need or will transform to MCP apps in order to be, to be rendered in Chat GPT.

Yep.

I wanted to bring you guys on here because I think what you're doing is very important. I'm also very impressed with how far along you came in there and now a part of the spec as well. And for many developers around the world, I think it's very important to highlight that. Like after, after an year of this protocol, there is continued building, continued, continued adoption, all of the major labs like adopting and working together, which is great. I absolutely love the fact that you guys are considering this not to be the next iOS versus Android Till now, developers who wanna build apps, they consider Hey, iOS, maybe first with one developer and then Android, is a different thing. The fragmentation o of the web, luckily on the web, HTML kind took over. And now like we develop apps for, browsers in one way. if and when. the whole world of AI is gonna come for us all, I think the standard is very much needed. So really a huge shout out to you guys for this incredible work. I know it's exciting to build demos and give talks, but the actual kind of like RFC process of standard is not as exciting, but somebody has to do it. And so I appreciate that. This is the two of you. Anything else that I haven't asked you yet that you absolutely wanna highlight? Or many other folks or anything else? This is your stage.

I think the main thing is again, go back to the community.

so this is very much an open standard and we are looking for more contributors and more opinions and obviously already dozens or hundreds of companies playing with it. But we need more just to make sure that we have everything covered. So we really invite you to go to CPI dev to the GitHub repo to the RFC and kind of get this, Going. Yeah. And I also reiterate how amazing this, time is to build things. just, everything is changing. The web is changing. ai, is building new environments, new runtime environments. So this is an amazing time to build things. just go ahead and build using MCP I using MCP apps or try to invent something new. But, this is the time.

So this is a demo.

Super quick, for CP ui, do dev that, you go in there, there's a chat with Qwen Q qm, cloud Force on it, and you can say something like show tasks and instead of just responding on the text, you can see that I have an interactive window with interactive bars, that I can like, just like work with. and this is like an within the chatbot. This is like a window that opens up and there's an interactivity in there. Obviously visually it looks completely different than something that, NLM would build. And this runs on a server. So this is like a quick demo of the work that you guys are doing. Thank you so much for coming up. We need to move on in the show. Yeah, I know that, you also have otherwise engagement as well, ido. And yeah, thank you so much for your work. we really appreciate you guys as you release more stuff, please don't hesitate to come back and tell us about this. and yeah, good luck setting up a standard for the new world of the, of LLM Internet. Cheers guys. Thank you. Thank you. Cool. Thanks Alex. Bye. we'll bring back Yam and Wolfram to cover super quick. the reason why all of this exists, the MCP UI, as we had the folks on the show before, is because the intelligence move forward, we're gonna cover Opus 4.5, as part of it as well. But can you guys believe it's only been one year of MCP with the crazy excitement and the M MCP peeling and then we were forgetting about this and started building again. And then the UI layer now adopted. This is honestly quite crazy. Yeah. Honestly, quite crazy

how we changed from model that just output text to

models that actually run comments. We had that before MCP, but MCP that's when it took off when you had so many MCP servers, you can integrate the tools and when OpenAI integrated it, I think that was a major milestone as well. Now you can use CPS in the big chat interfaces, Claude and tropic, and Tropic cloud. and

yeah.

Can you imagine that?

cloud 3.7, which I think many people recognize as a step, major step in AI coding capabilities was at least February.

and cloud code was dismay when was cloud code?

I think March or maybe also February, I think.

Yeah, something March. Yeah. Some way it's not even a year. and like now it's just the standard. Everyone is using some agent for something. wait. It wasn't like this a year ago. It's crazy. Absolutely crazy. And so

just as a visual example, Chat GPT, apps are now gonna move to

this like MCP apps, ui, sorry, spec. And here's an example of Zillow app that you like load within the Chat GPT interface. This is live for all of you as well. You just type Zillow, show me a map and shows you an interactive map. This is basically an iframe to Zillow. And you can see this is visually Zillow, IDO is right, like when, previously with MCP, maybe it will respond with some text and the Zillow brand will disappear. And this is not great for brands. And this is like a full app story. You can click and interact with it. And then the chat actually knows the context of what's going on in this app. also these are homes in New York, so don't look at the prices. but we have to move on because, we haven't covered the major labs and updates there. And for Opus 4.5 is a big one that we have to absolutely mention. So let's. Let's dive into Opus folks. Opus is the biggest brother of the Claude family. sonnet 4.5 obviously was one of the best coding models, maybe still, that there is. and how's Opus do you guys play with Opus? What are your thoughts about Opus? I would love to hear.

Okay.

I'll go second because I'll just waste all the time, but yeah, it's a mon it's a monster. It's a monster, absolutely. Yeah.

You are probably more qualified to talk about the coding

aspect, so I just chatted with it for a couple of hours actually. it'll refresh my, my, subscription, not the, I just open route actually, and the model is excellent. So it's amazing. I always tested in German as well, and the German is excellent. So from the big releases, Gemini is still my favorite, but it's so close now with the two. While was a bit disappointed, when you, when we talked about, the new 5.1, that was a bit disappointing for me. The releases, I don't dunno many people who use it, although it's so popular on Open router, that's strange, but now Clot is out and it's a big, the biggest model personally in my use, I did not see that much difference to Gemini or to Sonet even. So it's niche model, I think specifically for programming where it's probably the best, but it's also much more expensive than the other ones. So I think you use it for the most, difficult problems to make a plan. And then for instance, SONET can do it. I, yeah, but I'll pass to Yam you probably more into this.

So before Yama I should call out.

you mentioned the price. We should mention the Opus 4.5, the top tier coding intelligence that you can have access to right now. There was released this week in this week of Insanity where everybody released like the top tier models, one after another. And we kept beating state-of-the-art and Swyx bench every day on a daily basis. But, Opus 4.5 is three times as cheap as, the previous opus. So it costs around $5 for 1 million tokens, whereas the previous Opus cost 15, which means they probably have built breakthroughs in how to serve this model. so yeah. How's Opus for coding now that it's 80% of Swyx image verified in absolute state of the art and coding?

Okay, I'll start with what you don't hear on X, which

is not the best for every single task you can possibly think about, which I think it's important to say. this is from experience, couple of days of like daily driver heavy use. It is noticeably better than Sonet. There is no, you feel it on the first message. It's absolutely better than Sonet, but still, you get the impression on X that it's like everything you throw already, it doesn't matter. It's the best model for sure, for this task. That's not necessarily true in my experience. I do still use, codex, codex Max and Gemini, three for, they're good for different tasks. But what I wanna highlight, and, I said I'm keep saying it all the time because it is, because this is the main point I think that really differentiated the harness that it is. Use these powering cloud code in combination to how good Opus is. It makes the entire tool just irreplaceable. There is, it's the, as a full package. Both of them is nothing gets closed as an agent for the tasks themselves. I just wanna say for the tasks themselves, Opus is definitely, on, definitely on power, on many tasks. It is the best ish for sure. Opus knows a lot of tiny details. That you didn't know that you wanted the model even to know until you use it. And all of a sudden it just pulls tiny details about the stack, about the code, about things that it like, wait, that's like a really good idea. I didn't even know you can do that kind of feeling and you get, get it again and again. Moreover, what I really liked is that my problem with, my problem, my issues, you can, my main issue with cloud code on daily basis is that you run the thing and if it, that, if the task takes too long and you see the model goes again and more steps trying to solve the task, it loses track. And at one point you see the performance slowly deteriorate throughout the con the context and like it's forget what you want, even with the to-dos and so on. It like becoming, I don't know, everything becomes, messy throughout the context. And Opus just goes like the same phenomenon that you see with Codex on Worth mentioning that Codex feels like it can go forever, in my opinion. Code Codex is in a different level of going for long tasks, but Opus is definitely getting close to this. you can just let it. Do its own thing and watch the terminal hope for the best. And it does eventually do the task, which was not the case with sonnet at all. If you just let sonnet go for too long and you watch the terminal, you are going to watch it go in circles and do the same mistake. And still, again, you're absolutely right and it's not, yeah, it's not the same as Opus is definitely a step change from Sonet, tropic cooked. there is no, not no way to spin it otherwise they just cooked and we users enjoy the entire thing, the entire competition. We are the ones to enjoy. Like we get the best model every day. You get the best model. Yeah. Thank you. Thank you. Come on, keep competing. We get the best model every day. Yeah.

So on the previous segment I kinda mentioned, the extent of this year,

the 20, 25 year of agents, et cetera. I think the huge unlock for many, use cases was the tool use specifically how these models approach and use many tools in their thinking. And then as part of their execution and as part of the agenda run and flow as well. Opus 4.5 came out with like huge overhead cuts on many tool usage, 85% overhead cuts on like tool search. and then programmatic tool calling is something that they've announced with the MCP kinda UI spec where, the tools that require a lot of data usually. Before this kind of concept, Opus would have to load all the tools outputs. If you think about like tens of hundreds of Excel rows that you wanna manage, Opus would have to load this in the context to review, et cetera, to even call the next tool to say, Hey, these are the most important ones. So programmatic tool call, is now a thing where Opus essentially will run code to execute these tools within programmatic context if they're specified correctly in the MCP. And then based on this reduces like 37% of tokens on complex workflows as well. as always, one of the things to call out, and we call this out often the latest releases from these labs. What we see is because the age of kind of thinking again, started what a year and a in a few months ago. September was the first test time computer release from OpenAI. these models can achieve a lot of stuff with the thinking process. They can figure out docs, they can do a bunch of stuff, but this extends the context window and so they're getting lost, et cetera. more and more we see from these big labs, these big models, they think harder on longer problems on the actual problems, and then they think less on the very simple ones. And we see this kind of change when, instead of just saying, Hey, let's throw as much. Compute in the form of thinking tokens at a problem. They're now trying to figure out how to maximize and do the same hard difficulty problems with kind of less compute. And we see this with Opus as well, and just Atropic absolutely cooked. And to do this with a one third of the price of the previous Opus is very good for people. they also reduce kind of the Opus usage on Cloud Opus Max, I believe. yeah, very impressive release, community. absolutely. We have some community reactions here in our helpful infographic. community built 3D simulators and, 3000 line refactor with Cloud Opus. Artificial analysis puts, cloud Opus 4.5, at number two, overall intelligence index, 11 points over prior Opus, and I think it's just behind Gemini. it still costs six times as much as Gemini on some emails, right? So this is like still a consideration for you as well. And, oh, and we should mention they mentioned specifically Excel. So specifically Excel was called out in their kind of like release that this model is like really good at, like helping you do some very crazy Excel tasks. I don't use Excel often or at all. very happy for this. Very thankful in Thanksgiving that I don't need to use Excel, but I know that many people do and they like it and like they need some AI help and Opus is there to help them. So this is an update from Opus, Tropic absolutely cooked. and, you guys should definitely try it out in cloud code. I think it's up in corer and windsurf and like all of the other ones besides Antigravity, which is Google one. And, it's not there. anything else left to say on open? Let's see if we covered everything. just incredible model overall Anthropic. Cooked. shall we move on? I think we have a little bit of time for some video stuff. So maybe we'll just shout out the video stuff from super quick. the two models that were released in video was Nia on video, 4.5, which we told you about, and then LTX released retake, which is really cool. Retake is the AI video.

I tested it.

Actually, I was not as impressed in my test, but I just did one test. what it does is you can upload a video up to eight seconds, I think, and then you give it a prompt and it'll recreate the scene in the way you prompt it. Even change the audio, though, that sounds extremely impressive on paper. And I think there some demos that show it very well.

Yeah.

So here's the demo of an astronaut going to the moon, just like doing a back flip or exposing the, the, the landing was basically, a Hollywood production. And, putting some other folks, so basically LTX rewrites these, videos. I think when, SOA was announced, they basically announced that this is the thing that they support as well, isn't it? the, you can like re-edit videos or remix videos. but, LTX Studio, just to mention, LTX Studio is an actual studio for video production. And so not only the releasing models, and I think this model is not, in open source. they're giving you an end-to-end set of tools for a video editing. So it would make sense for this thing to exist. And the examples are pretty cool. let's see, like this, so you can rephrase dialogue like Wolfram, like you said. And then, you can see visual changes in, in how, this video works and the character consists is pretty cool. And then the, go ahead.

There's a lot of potential for funny clips.

what I tried to do in my demo was, to take the scene where Hanzo shoots, the bounty hunter and I changed it to the Bounty Hunter, which shoot Hanzo nice. He shot first. who shot first?

Who shot first.

This is the age old question.

Created different alien and itself.

So it was funny, but it wasn't what I wanted.

So here's another example on replicate, and this is live on Replicate.

this is the original video, and what they did is, I'm not sure what they did here, remove the lens flare. Oh, wow. Okay. So this is pretty cool. This is like a very small thing, but you can look on the left. here you have a lens flare for this model. And on the right, there is no lens anymore on this model. which is pretty cool. And then the consistency is fairly, pretty much kept the same. So it's the same person, or at least it looks like the same person. Like it didn't screw up, it just edited the lens fla out. So like in this shot, let's say here, and then she has lens flair at this shot. when it like goes around here and there's no LENR here. Yeah. So it actually worked very well. Very impressive. very specific use case.

Yeah.

What we are seeing is the image editing moment for video. So you can take no Photoshop for video and change it the way you want it because we all know it's easy to get a creation with AI that, is about 80% where you want to go. But if you really want to nail it, you want a hundred percent, you have to edit it somehow. Usually just recreate, change the prompt or something. But in this way you could take a video which takes long to create and is expensive and just recreate the part that you want changed. So that is a big step ahead for video creation, I think.

A hundred percent.

And speaking of big step ahead of video creation, the last news that we have for you is a new. Stealth model called Whisper Thunder is number one on ELO score on the video arena from a, I think this is, artificial analysis tops VO three and tops click 2.5. this arena is with no sound and, it has a highest ELO score out of all of them. And the speculations are it, we don't know which video this is, but this is the stuff that it generates. Let me, lemme play for you. it looks absolutely insane and this is compressed video, but, like the prompt adherence is crazy and the actual video support is crazy. the new model and once we know what the model is, we will definitely tell you with this folks, I think we're at exactly two hours in our show for Thanksgiving. So happy Thanksgiving to all of you.

I wanted to say maybe we could take the opportunity,

think as Thanksgiving and, decide what we want to say. Thanks for.

Yes, let's do it a hundred percent.

And we also will invite the folks in the comments also to do so with us as well. So since you proposed it, how about you go first?

Okay.

So I want to thank everybody who is contributing through the wonderful community of ai, the people that release the models that create open source stuff or create tutorials or just share their knowledge. And, that includes you, Alex, as well, and you yum as well. on, and this of course, so everybody who participates and is not just doing it for the money, but we're doing it because we love the community. We love to learn and to contribute. And, I wanna thank everybody who is in this. Thank you.

Beautiful.

Thank you. Thank you, Wolfram. Wait, yum. Your turn.

I just wanna thank that.

I just wanna thank the fact that we are at the point where there is no, there are no, competition between open source, closed source at this point. It's feel like, everything is, is moving on track and, for the, for all the people that's working hard, making all of this happen, doesn't matter, which I strongly believe that even if you posted some random code somewhere that does something and you think, ah, nobody's even looking at it. the internet is large. Chances are if someone saw it and use, the kind of thing that you work hard for, even if it feels like not. So for everyone working hard and, I don't know, helping us accelerate towards. Whatever we are going to, which I'm not sure what you, but whatever we're going to pretty much,

that's beautiful.

And I specifically wanna thank, I wanna thank Ashish Ashwani, no, Niki Parmar, Jacob, Leon, James, Aiden Gomez, Lucas Kaiser, and Ilya Polosukhin for the 2017 favor. Attention is all we need to bring us.

And for Jensen, also Jensen for the G. Yeah.

So for attention

is all we need.

And then I wanna thank Jensen for the, to, for the scale up of the insane, GPU infrastructure that we have for this AI intelligence boom that we are here to cover. I wanna thank Wolfram, Nisten, yam, LDJ, Ryan Carson, all of the Thursday, I co-host and guests as well. the come and make the show. I wanna thank you the audience for tuning in with us. Even on Thanksgiving, we have like over a thousand folks, joining us on this Thanksgiving for the ability, for us to do what we love, which is keep up to date with everything that happens in the world, the way I try out all the tools and then teach each other and tell you as well, and learn from you as well. A bunch of the things that we learn about come from the community in the comments. So I really wanna thank everybody here on the 2025, Thanksgiving. On Thursday. I'll also say that if you go back to 2024 Thanksgiving episode and the previous one, you will be shocked at how far we've come along in the world of ai. Just shocked. And we have this thing called hedonic adaptation where we get new tools, we get super excited and like super, super quickly we move on. for example. Nana, banana, broken General, these like awesome infographics that we showed you before. And, GPT image that barely anybody uses now could turn everybody into Ghibli. And the Gib Fi moment was like this huge thing. And now we're like, oh, of course you can turn every image into a Ghibli character. This was not a thing a year ago and this is like a thing now. And I just wanna appreciate how fast these models, the intelligence moves and appreciate also the fact that, we should stay humble. We should celebrate these releases. We should understand how amazing of a world we're living in now, where all of these powers are available. And, we are just very happy to be here and have the opportunity to be here to tell you about this, and then also celebrate them as well. And the last, my thanks goes out to Apache two as a concept. And MIT as well. And all of the labs that release open source tools, that other people can iterate on, build on and improve, basically, tide the raises all boats. So for this thank you all also for joining us for this Thanksgiving, 2025 edition of Thursday. I, if you missed any part of the show, please subscribe to ThursdAI news where you basically get. All of the news that we've covered, all of the links as well, an edited version of this live podcast there with subtitles as well. pre subscribe on YouTube as well so you never miss a part of the show. Thank you so much and we'll see you here next week and, I'm sure they're gonna be amazing releases by then as well with this. Thank you all. Bye-bye. Have fun on Thanksgiving. Enjoy.