Limitless Podcast - Kimi K2.5: The Best New Model is Open-Source (Again!)

Episode Date: January 29, 2026

Kimi K2.5 from Moonshot Labs, live now, employs multimodal training to process 15 trillion tokens from various formats. This model allows users to create website replicas from screen recordings in moments, drastically reducing operational costs to $0.60 per million tokens. We discuss Kimi K2.5’s efficiency in handling complex tasks with up to 100 sub-agents and its implications for open-source AI.------🌌 LIMITLESS HQ ⬇️NEWSLETTER:    https://limitlessft.substack.com/FOLLOW ON X:   https://x.com/LimitlessFTSPOTIFY:             https://open.spotify.com/show/5oV29YUL8AzzwXkxEXlRMQAPPLE:                 https://podcasts.apple.com/us/podcast/limitless-podcast/id1813210890RSS FEED:           https://limitlessft.substack.com/------TIMESTAMPS0:00 The Rise of AI Agents1:43 The Game-Changer: Kimi K2.53:40 The Power of Sub-Agents5:26 Efficiency in AI Tasks8:40 The Role of the Orchestrator10:13 Creative Applications of Kimi 2.511:54 Cost Comparisons of AI Models16:03 Strategies Behind Competitive Pricing17:54 The Genius Behind the Model20:15 Open Source vs. Frontier Models22:36 The Future of AI Development24:31 Engaging with AI Tools------RESOURCESJosh: https://x.com/JoshKaleEjaaz: https://x.com/cryptopunk7213------Not financial or tax advice. See our investment disclosures here:https://www.bankless.com/disclosures⁠

Over the last few weeks, one of the hottest new topics that exist in the world of AI has been agents, agents that can actually get into your computer and do things for you. We've seen this with Claude Co-work, which can actually access files and make changes on your computer. And then to the fullest extent, we just recorded an episode on Claude Bob, which allows an actual computer to fully take over your life, send messages on your behalf, take care of emails, book reservations. The problem with that is that even an afternoon of use can cost hundreds of dollars. So what we've done today is we've actually figured out a way to replace that totally for free,

where you get the same quality outputs but with none of the cost. And to do that, I'm going to start on Anthropics website because the reality is that the screen that you're seeing right now isn't actually Anthropics website. In fact, it was built using this new tool completely for free in about 25 minutes, which I thought was such an amazing demo. It was built using a tool called Kimi K2.5, which is the newest model coming out of China that is fully open source, fully open weight. And in order to build this, all I had to do was feed it a video.

So you'll see on the screen here, I generated a video on my desktop using a screen recorder that copied the Anthropic website. I said, hey, just take the screen recording of the website and create an exact replica of the website for me. 25 minutes later, without any additional prompts, it listed all of the things it did. It went through all of the design. And then actually published a full preview of the website that we can see here on this read record.

So this model is incredible. I don't know if you've had a chance to play around with it or check it out. but this is a huge change in the world of agents because of how capable it is for such a low cost. This model is trending pretty heavily online right now. I actually saw someone describe Moonshot Labs, the creator of this model, as the Anthropic of China. This was a quiet release, Josh. So the creators of this model kind of just updated their chatbot interface with Kimi K2.5 and didn't tell anyone. And within a few hours of that launch, remember, no publicist or anything, it was the number one trending model on Hugging Face, which is like where everyone goes to access all these free open source AI models.

And what you just demonstrated, I think, is one of the core reasons why this model is so special. So to give a few stats about this, it was trained on about 15 trillion tokens. And typically AI models are trained on text tokens specifically. This wasn't the case for KimiK-2.5. It was trained on text and audio and visual and a bunch of other mediums. And the reason why this is important is it allows you to do the example that you just show Josh, which is feed it a video or in this case a screen recording of a website that you wanted to build and build it in exactly that way.

And the reason why this is important is it shifts the use of AI models from explaining what you need to do to it. So like, say like, hey, like could you do this, like describing what you want from it into just showing it what you want to build. And I think that that's like a really intuitive way for people to interact with AI models versus like people that aren't just quite literal like me sometimes when I'm trying to explain something, right? The second really cool thing is you started off this episode mentioning agents, Josh. And I think this is really important because Kimmy K2.5 has this superpower where they can

spin up up to a hundred sub-agents. Think of a sub-agent as just another instance or replica of Kimi K2.5. but it's specifically focused on doing a certain task. So for example, if your goal is to figure out whether investing in Anthropic is a good idea, it'll spin up one agent that does the research, another agent that does the fact-checking, another that tests different kind of architectures. And the cool part about this is they can work in parallel, which means that you can cut down the execution time for a task by four and a half times.

So imagine you had a task that took four and a half hours. you can now do it in one hour. And I think this kind of like multi-agent trend that you identified or that you spoke about is super important because that's what we're seeing with the likes of Anthropic with Cloud Code and and Open AI with Codex. But the fact that this thing is free is completely insane. Josh, do you know how much it cost rumored for them to train this? I have no idea, but I would imagine a tremendous amount of money for them to just train it

and then release it fully open source, open weight. So the rumor, and again, this is not a fact. I wish I could fact-check this. And also, to be fair, before you say this, the Chinese models are notorious for lying about how much it costs. Yes, they are. So take this number with the grain of salt. You're right.

So the number that's being floated is $4.6 million, which is nothing. It seems so low, which is nothing compared to the billions of dollars that OpenAIA has spent to kind of train their models. And to give you guys an idea, like why we're comparing Kimmy Kade, 2.5 to these like frontier AI models built by Open A&A and Anthropic is because in some cases it's almost as good as this. Like if you look at its performance on humanity's last exam, which is notoriously the hardest benchmark for an AI model to be tested on, it scored a 50.2%

which beats Claude's latest model, Opus 4.5 and GPD 5.2. It doesn't quite beat Anthropic at coding, Josh. I know you built that cool website in a few minutes, which makes you think that maybe it's really good at front-end development, but just a really impressive model. And I'm guessing it's like super cheap to operate as well compared to like some of these expensive models. Yeah, we're going to get into the cost because if you do want to use it at length and you don't have a couple H100 GPU sitting in your home, you're going to have to pay a little bit. Thankfully, it's significantly less. And we'll get into the prices. But one thing you mentioned is that it's actually not the best coding model in the world. And I think that's okay. That's not the real

breakthrough. One of the most amazing breakthroughs is actually before we were recording the show, Eajas, you showed a demo of you gave Claude code the website of Figma and said, hey, can you go emulate this website? And it actually did a pretty good job. The difference between Claude Code and something like this new model is that I was able to feed it just a video. And what it did is it analyzed each frame, each pixel within each frame, understood the context of each pixel, and then figured out how to intuitively regenerate that in a web page using code and like whatever type of design tools that it used. And that is the novel thing, because most models do image to code, but Kimmy K2.5 does video to understanding to code. And I think that's one of the more novel breakthroughs.

One is the three, actually. The second, with it being natively multimodal, so you mentioned 15 trillion tokens that it was trained on, but that's mixed between visuals and text for the first time. So this really has a good understanding of videos, of photos. It's starting to even get the Google physics. And then the third part that you mentioned, which is the agent swarm. I want to spend a little bit time on this because the Asian swarms are super cool. We actually have a example of one of the Asian swarms and how it works. The way it works is, it's it's able to separate itself into basically a hundred small mini-taskes. And the example that we're seeing on screen now is a film script. And it's a short story that the model generated. It created a

shot list. It created renderings of images of what the frames of this could look. And it generated basically an entire movie in a fraction of the time that it would take to do as a single model. I think the actual number is four and a half times faster. Like you mentioned earlier versus traditional models. So it can call up to 1,500 tools. It's like this swarm of agents working on a single problem. So it's faster. It's more efficient. It can, it's just like, you can make a movie script in five minutes. And it'll generate the entire thing for you with the shotless and all. I mean, some of these examples are pretty unbelievable. Did you hear about the underlying mechanism that they use to build this? It's actually super cool because, well, one thing Chinese

AI model labs are repeatedly known for doing is, you know, they don't get access to all these fun, expensive GPUs that the Western labs get to. So they had to get really creative in their research and training techniques. And they did that with the agents. So to get that four and a half times efficiency that you mentioned, they used this technique called parallel agent reinforcement learning. Typically, when you spin up 100 agents, you're going to have a hard time. And the reason why you're going to have a hard time is something called agent collapse.

So typically a model will be used to doing things in sequence. So if you ask it to do a really complex task, it's going to start with task one and only proceed to task two once it's done with task one. And if you spend up a bunch of agents, the model might sometimes still do things sequentially. And you don't want that. You want it to do in parallel. So this new training technique that they spun up and pioneered, there's a paper all about this, is super unique and never been done before, which allows them to not get model collapse at 100 agents. the crazier part is each of these agents get access to over 1,500 tools.

And that's what makes an agent useful. Like you go from an LLM telling you what is useful to an agent that can actually do something on your behalf. That's pretty impressive. And then the final thing is they have this thing called, well, I actually don't know what it's called technically, but it's the way I understood it is it's kind of like a brain. It's called an orchestrator. And so it ingests the tasks that you've asked it to do. And it breaks it down into multiple different tasks.

the fact that it could do it for this cheap, Josh, sorry, I know I keep mentioning the cost, but you need to tell all the people the cost because it's just insane. Like, this is something that I would use regularly, a company would use regularly, because it saves them so much money. And we actually have this really cool visual of the orchestrator on screen here,

which gives you a visual representation of what that looks like. So the orchestrator breaks this down into sub-agents. It assigns them tasks, and then the task kind of go back and forth through a fact-checker. There's a file downloader, there's a web developer, there's this entire toolkit. it. So one of these is emulating an AI researcher, the other is a physics researcher, the other is a life science researcher. And what you're getting is a series of experts across every

domain working on problems in parallel with access to 1,500 of these tools like the fact checker, the file downloaded, the web developer, the text scraper so it can view images and understand and interpret what they mean. And it's such a powerful stack that you have. And without the collapse, with this software novelty that they've introduced, it allows them to do this unbelievable thing. So to your point, when, I mean, historically China has been hardware constrained and they've really accelerated on the software. And this is very much an extension of that acceleration. Now, we have some more examples that are very fun to show that. I would love to show because as I was going through to prepare for the episode this morning, I was like, wow, this is pretty cool. And I jazz,

you even dropped in one of your own that I thought was pretty neat. So if you don't mind explaining what this one is here. Well, it may not be the coolest example, but this is something that I would personally do. And I know a bunch of my friends would do in their spare time, which is just again, on websites, the fidelity of these things is pretty insane. And I have to emphasize we're going from a screen recording to like a fully functional front-end development. And I don't think people quite understand how necessarily hard it is to do front-end development. I think a lot of software engineers this will kind of scoff at that comment.

But it is true. It is like super hard to do because there's the design element, which is incredibly subjective, which is Kimmy K2.5's exact point. Instead of trying to describe it to an LLM, you can just kind of take a screen recording and spin this up in a matter of minutes, right? It took you, I think, 7.5 minutes. I just want to like emphasize a point that you made earlier before this, Josh, which is the agent side of this model is super important because if you look at a model from Anthropic, their flagship product is called Co-Work and, sorry, called Code and recently called Co-Work. If I were to like tie both of those products, in one unique trait. It's the fact that you can spin up multiple agents. In fact, the founder of Claude Code, and in fact, a lot of the Anthropic team

do between 80 to 100% of production level code. So that means new products that they ship completely built by Claude Code. Now, they're not doing this using one model. They're doing this spinning up several instances of Claude Code. So to kind of like put this into perspective, this is the future of software development.

And software development pretty much under any major breakthrough for any industry going forward. Software and tech underpins everything. So if you have an AI model that costs a fraction of the amount that the frontier flagship model from Anthropic does and is 100% free in open source, yeah, you might need whatever, 50K to 100Ks worth of GPUs to run it on your own instance, but you can get access to Kimi K2.5's API right now. That is a huge advantage. And the Chinese AI labs just somehow stay on top. I don't know how they do this. Yeah, well, if we're talking into about token price,

maybe we could get into the economics first. We'll skip ahead a little bit because I think the economics of this is super interesting where if you don't have access to the GPUs in your house, which I'm guessing nobody listening to this episode really does. If you don't work at a major AI lab, well, there are ways in which you can run this for free or close to free. Now, the example that I showed earlier where I created a website clone, that was free because Kimmy gives you three agentic tasks per week,

basically, that you can use for free. But after you've exhausted that, there are some economic pricing sheets that we can use to kind of compare this to other models in terms of cost. So for Opus 4.5, which is the most popular model that we've been using, everyone's been using, it's fairly expensive. The price input per token, per million tokens, is $5, while the output is $25. So for every million tokens you generate with Opus 4.5, which is the flagship coding model, it costs $25. For Kimmy K2.5, the input is 60 cents per million and the output is $3.00. That is almost a full order of magnitude, 90% decrease in price relative to Opus 4.5 at a very comparable rate. And that's just if you're comparing it on

code. If you're comparing it on general agenetic tasks, it's actually slightly more capable than Opus 4.5 for one-tenth of the cost. So if you're using something like Cloudbot, which we recorded an amazing episode on earlier this week, which you should go check it out, you can just swap in this new model and run it through whatever cloud service you want. And the price of your agent will be one-tenth of the cost. And this happened in the matter of a couple of of days. So the costs are rapidly decreasing. And I think that advantage of it, one being open source, but two being cost-effective, is huge for everyone. I just if you remember, I think it was two or three weeks ago, Anthropic cut off XAI from using Claude Code. They actually removed access to it.

And because it's closed source, there's nothing they could do about it. But if they're using a model like Kimi K2 to run K2.5, to run their agents to build their code, there's no one to who can actually sever that high. And it's the same for developers, where if you're building on a platform and you don't want it to change, well, now you have the open weights. It's going to be locked in forever. It's going to be a fraction of the cost. This is a really viable substitute for those who are loving the agentic life OS workflows.

Do you have any idea how they're able to produce this for such cheap costs? Because like, I'm trying to rack my brain around this, right? So, okay, sure, you've made a few research breakthroughs. the Chinese labs in particular are known for discovering or commercializing mixture of experts, which kind of cut down prompting and inference costs and training in general to a fraction of the price. But still, like they don't have access to some of the top hardware, right? And kind of like Moore's Law would state that eventually a bunch of these GPUs that are A grade are going to cost so much less and run like 100x more inference performance per token.

so it's going to cost a lot less in general over time. They don't have access to these resources that the West does, right? So you mentioned like Anthropics cost, right? You said what? It was like $5 in and how much out? $25 per million tokens compared to three. Okay, that used to be $15 in and $75 out when they launched the product. So we've come down by a significant factor since then.

But again, I would imagine they did this because of cheaper, more effective chips. How is Kimmy K2.5 done that? How has Moonshot done that? I think I have two answers to this question. The first is through software innovation. I assume they have cracked some sort of a code that allows them to generate less tokens per output. The second one is the margins. Ejas, how much money did Anthropic make last year?

It was, what, $10 billion of revenue? $10 billion. Correct. China is unfortunately not the leader in AI, and therefore they need every incentive in the world to dethrone the leader of AI. One of the ways you could do that is by winning out the margins and cutting down those margins for your competitors. clearly they have this strategy because they're publishing this open source, open weight,

and you're seeing that happen with the pricing as well. I assume a large part of that revenue from Anthropic is just margin on the inference that they're charging. It doesn't cost them anywhere near $25 per million tokens, but they're able to charge for it because they're the leading frontier model that all of these labs and businesses are willing to pay in order to use their services. In the case of Kimmy K2.5, they don't care. They don't need to make profit.

They just want them in market share. And to do that, they're able to undercut pretty aggressively here. Like, I'm sure Anthropic could match this and perhaps not actually lose money, but that profit thing is real. It also helps that they have an absolute gigabrain as their founder and CEO. I don't know if you've looked into this guy, but this dude is only 31 years old. He was born in China. He went to Singhua University, which is actually the most popular university for AI and

ML researchers in the world to graduate from, 50% of. of the world's top AI researchers, by the way, reside in China, and a large chunk of them graduated from Tsinghua. But Josh, he also did his PhD at Carnegie Mellon, and he did it in under four years in, assumedly robotics and machine learning, which is very impressive. And he also did a very long stint building out Google Brain and meta-AI research.

So he was probably one of those meta-researchers getting paid tens of millions of dollars here. So this guy's track record is insane. So it doesn't, I guess, with that CV, doesn't surprise me that he's made these breakthroughs somehow, even on the hardware that he's constrained on. Yeah, it's incredibly impressive. I'd love to hear more from them. In fact, we actually, the first time we heard from the founder was earlier today with the announcement post. I had never really seen what he looked like. I hadn't really heard him communicate. It feels like it's a very sheltered, kind of quiet, secretive workplace that they have there. But I'm hopeful that we'll

start to see more because, my God, the talent there must be unbelievably impressive. Just in China in general. When we talk a lot about the trading competitions that we have, China's always seemingly winning, they're doing really well and clearly they have incredible talent density. Now, you're showing on screen something that I'm very excited to talk about, which is jumping back to the examples of what you can actually do with this new model. And one of them is this really fun blueprint to 3D model space. Now, Ijaz, you've watched friends, right? This might look familiar to you. Oh, yeah. Mm-hmm. Not the one on the left, yeah, not the exact high fidelity. You don't know the blueprint of those spaces, the sets.

Yeah, it's pretty cool. So it took a two-dimensional blueprint of a room, and it generated a three-dimensional version of Monica's apartment, or Monica and Rachel's apartment. I haven't watched Friends, but I know it's very popular. And I've seen clips from this room, so I'm familiar which one it is. And it's a testament to the types of new creative things that you can do now that it has the image to critical thinking to output type of thinking process through generating these outputs. And I just thought that was really interesting. There's a lot of really fun use cases that you can

use. Dude, this is a this is a $10,000 a month apartment at minimum, Josh. It's making me feel poor looking at the schematic. Oh my God. Yeah, right. Grove Street, New York City, that might even be more than 10 grand. That's prime real estate. Come on, dude. Yeah, that's insane. How are they able to pay rent? They were making comedic jokes the entire time for seven years. But also, this becomes a very useful tool for real estate agents, right? Because they want to kind of recreate spaces, allow you to feel and live in the space more. And granted, this is a low fidelity version, but I'm sure this is step one in creating some higher fidelity mockups of spaces that you would possibly want to rent. If you're building a house, if you're building anything, this is great for construction,

for modeling. These services used to cost a ton of money for virtual renderings. Now they're effectively free or very close to it, maybe just a couple cents per output. And that decrease is pretty It's substantial. Open source is having quite the week. They're having a moment. I've commented a lot about this before, but I've said that I never think open source will actually ever catch up to frontier level capabilities. And in this case, in some ways it does, in some ways it doesn't.

Josh, you know, in my period or era of life right now, I am a coding agent maxi. I'm incredibly bullish on Anthropic. So, you know, I scrutinize any other competitor pretty heavily when it comes down to this. I don't think it is as good as code. You mentioned this earlier, but it's scarily good in some aspects, right, with the front end development.

So I'm curious to see how people use this. And I think what I love most about this is a lot of my friends that kind of want to do more creative pursuits, like build websites and do more front end stuff. They don't want to pay 200 bucks a month, Claude Code max, right? But they can get this for free and they can access it today. You literally built your website today,

like in a few minutes before the show started and recorded it. 25 minutes with one prompt. That's insane. That's insane. So if you can do it, if I can do it, anyone else listen to this can do it. Definitely go give it a go. Like I want to see some examples that people kind of like do with this. The Kimmy website itself actually has a bunch of ideas and use cases that you can use to kind of emulate or get inspired by. And this is one of the major things with the model launch. Like the reason we're talking about this today is because they provide this really awesome demo online of them screen recording a website and then emulating that and creating it in five minutes. what we did, and that's why it's so exciting. So the models that are not only able to make it accessible through lower cost pricing, but to kind of give you these curated experiences

where you can satisfy some sort of goal that you want in a way that's easy. All I asked was, hey, just create a clone of this website, do it identically and don't make any mistakes, and it did it in one shot. I think that is a critical threshold required to onboard a lot more people to be excited to use this stuff. To go and set up quadbot, it's pretty technically challenging. It takes a little while. It's not for the faint of heart. But something like this where they give you these use cases, they make it available for basically free Miam, where you can pay extra if you want to use it more. It's really exciting to see. And yeah, I mean, China's totally having a moment. And open source is totally having a moment. Like both of these things are converging

at once to create all of the news this week, while the major AI labs who are closed source are just kind of working in silence, perhaps trying to figure out how to best react to something like this that becomes open source and available. Now, you have to imagine, EJAS, it's been a little while since we got a new a new big dog on the block, a new frontier model for these labs. So the silence is deafening, but generally the longer the silence goes, the bigger the boom that follows. And I suspect we are only a few weeks away from some new models that will make this Kimmy K2.5 look like child's play, which is crazy to see because right now it feels unbelievable and magical, but I'm sure it is soon to be dethroned when the new models come out.

So I guess we'll be here to follow along with all of that news. Any final thoughts before we part today? I mean, once again, the clear winner for all of this is the users. Big time. We get access to all these frontier models for either a cheap or free option. It's super cool. Or if you want to pay the extra amount and get like a curated experience, you can also do that as well. If you want to use a Chinese AI model, go for it.

If you want to use a Western Lab, choose, you know, pick your poison. What I'll finish up with is the pace of development for these things, Josh, is so underrated. Like, I feel like we are so spoiled. When we first started this show around like eight months ago, we were like, oh, man, like, it can produce a pretty good market summary of this investment, but like, it's nothing like crazy. Fast forward to today, and I'm reading a tweet on my timeline from the founder of Claude saying, like, yeah, 100% of the code that we make, aka every new product that we build going forward, is managed by Claude, like is managed by Anthropic.

And I can assume that where the product like Kimmy K2.2.4.2.4. five, they're probably doing the same thing. So are we entering the era where AI just builds itself, probably, super scary. I read an essay last night, word to the wise, don't read scary essays at night, where Dario Omode, founder of Anthropic, wrote about his bearish thesis and why we need to be super careful going forwards because we're entering AGI, dare I say? I don't know. I'm super excited.

These model developments are super cool. And I'm excited for Josh, Codex. is probably going to come up with an upgrade. Open AI's new coding model is coming on in the next couple weeks. I'm excited they're having a town hall today. Fingers crossed, they probably want to announce it, but maybe. But when they do, this will be the first platform to hear it on.

Now, I know a bunch of you have listened to this and are thinking, hmm, I'm going to download Kimmy K2.5 or just use it and test it out. I have a task for you to try out. In fact, it involves not one, but two sub-agents. Number one, ask it what the top AI show is on YouTube. or any favorite platform that you listen to or hear on. And then ask it to subscribe if you aren't. Turn on notifications and give it a five-star rating.

I have asked this of you for the Claude Clorbot episode. I'm going to ask it for you for any of the Kimmy K-2.5 fans out. Please support us. It helps us massively. Yeah. If you ever need a use case, you can just have it. Go figure out how to subscribe autonomously to the YouTube channel. That'd be pretty cool.

And share it with your 10 closest friends through I message once you get it hooked up with Claudebot. That would be great. But yes, all of these cool, exciting new things that you were talking about, including Dario's Anthropic Letter and the Open AI State of the Union that they're kind of hosting today. We're going to cover that on our episode later this week in the AI Roundup. So stay tuned for that. And yeah, we'll see you guys in that episode. Thanks for watching. See you.