Build Anything with Kimi K3, Here’s How

My name is David Andre and here is everything you need to know about Kim K3. So first of all, Kim K3 is insane. This is a new AI model from the Chinese company Moonshot AI and it is by far the most intelligent open source model in the world. Even matching the power of Fable 5 and GBT56 Soul on many benchmarks and on some of them it's even better. For example, if we look at the front-end code arena, we can see that Kim K3 is the first time ever where an open source model is better than any other closed source model in the world. And this really is the central battle, open versus closed. You have to realize the release of Kim K3 is being positioned as China versus America. But that is a complete distraction. In reality, the real battle is between open source AI and closed source AI. And this is the first time ever in the history of AI where we have an open- source model crushing and beating the best closed source models out there. I mean, just look at these benchmarks. Kim K3 beating GBD 5.5 almost to the level of Fable. Frontier SWE better than 5.6 Soul. On Terminal Bench, it's better than Fable. On Program Bench, it's better than any other model. And same with S. Marapon, where it beats both Fable and GPD 5.6 Soul. So, in this video, I will explain why the release of Kim K3 is a historic moment. We will benchmark it against Fable 5 and GBD 5.6 Soul head-to-head, and I'll show you how you can start using Kim K3 today in your favorite agent harness that includes cloth code and CEX. Oh, and by the way, if you want to do that right now and you want to begin using Kim Kree in your favorite agent harness, I created a skill that will do the entire setup for you. It's going to be available as the second link below the video. Again, it's completely free, so go grab it now. Now, as I mentioned, Kim is the best open source model, and a big part of that is because it is the biggest open source model in terms of the parameters. In fact, it's almost twice as big as Deepseek V4 Pro, which previously was one of the best open source models in the world. Kim K3 lands at enormous 2.8 trillion parameters. So, when I saw this model drop, I knew I had to make a video on it. So, I asked the team behind Kimmy to sponsor this video, and they agreed. So, if you want to build on top of Kim K3, make sure to use the Kimmy API. It's going to be linked below the video as the first link. And when you purchase your first batch of API credits, you will receive an additional 15% completely for free. And this applies to any new users of the Kimmy platform. So again, if you want to build on top of Kim Kree, one of the best AI models in the world, make sure to click the first link below video and get started today. Okay. So, do you use Kimmy or do you use Fable? This is the question that a lot of people will begin asking. And actually, the answer is you should be using both. Fireworks AI did a great research on over 1,000 different agentic tasks and they figured that if you route it per task, the best performance is always the routing, right? So for tasks where Kim Kfrey is better, like front end, like 3D, like legal, you want to use Kimmy. But for tasks where Fable is still better, you want to use Fable. But the best possible answer for any type of task is use the model that performs best on evals and benchmarks for that specific task. say like, "Okay, I'm doing legal work now. I'm going to use Kim K3." Because if you've seen the charts from Harvey, Kim K3 is insanely good at any legal questions or consultations. So, do not even consider using Fable or 5.6 Sol. Just use Kimmy for those types of tasks. Now, if you're wondering what Kimmy can do on front end, well, look at this. This is Mac OS, an entire operating system built by Kimmy. And yes, it works. You can see like the folders work, you know, scroll bar, list, view. This is this is amazing. In the past, we tested the models on some oneshots and, you know, simple front end designs. Now, they're building entire operating systems for us to measure which model is better. Really, what a time to be alive. And to see the kind of impact Kimmy has, only look at the actions of other competing AI labs. For example, Enthropic realized that if they remove Fable from the plans, they're going to be screwed, right? Open release GPD 5.6 Soul. Kim K3 is matching Fable on a lot of the tasks for a fraction of the cost. So they realized, okay, if we remove it because they were threatening every single week and I called them out like what are these childish games, you [clears throat] know, oh we're extending another week, oh then it's going to get removed. No, we're actually keeping it. It's like either remove it or keep it. But they realized that they really need to keep it because of the competition. And I think the real reason why OpenAI Anthropic are really shivering and really taking this seriously is because of the pricing. Kim K3 is currently three times cheaper than Fable while being of comparable quality. You know, matching it on many benchmarks, even beating it on some for three times less. But what people are missing is that Kim K3 is an open source model, which means once the weights, the model weights are available publicly, which is going to be happening on the 27th by the way, there's instantly within the next few days going to be 20 plus different inference providers offering Kim K3. Some of them are going to run fast inference for a bit more. Some of them are going to run slower inference for very cheap. And that cannot happen for closed source models that are only hosted by a select few providers. Now, when talking about pricing, everybody looks at input tokens, output tokens. And while that is important, what is far more important is the actual cost per task. Let me show you an example. Deepswe, it's becoming one of the most famous and popular benchmarks right now in AI. And we can look at something like Sonet 5. You can see that it's $2, $10, right? It's a medium price model. Not the most expensive, but definitely not cheap either, right? So, you would think, okay, this is definitely going to be cheaper than something like Fable, which is probably the most expensive model in the world right now, right? 1050, literally five, times more expensive. But actually, what matters is the cost per task. If we look at Deepswe, we can see that Sonet 5 is the single most expensive model right here. Why? Because to complete the same task, which in this case is this benchmark on 54% success rate, it requires way more runs, way more turns, right? You might take somebody like Elon Musk and you can tell him build a million dollar company. It's going to take him one phone call, he'll do it in literally less than an hour to start a new million-doll company. But if you give that task to the average person on the stream, they cannot do it in their whole lifetime. This is the actual pricing that matters. Cost per task. Here's another example, but instead of comparing Kim K3 against Solid 5, which is a much worse model, we're going to compare to Fable, which is, you know, model of similar quality. Now, for 3D, you can see that this is clearly better, right? This is London. Kim K3 has more detail. Um, the river is nicer, the buildings are nicer than Fable. And also, it did it for less than half the cost. So, again, what matters is the task. The task was completed and Kimmy cost a lot less than Fable. So I already showed you that Kim K3 is the best at front- end design and front end coding in the world. But it's also by far the best model at 3D design. So 3D modeling, 3D objects, anything regarding 3D hugely incredible what the Kimmy team did on front end, on legal and on design. Just take a quick look. Okay, here is example of the type of front end that Kimmy was able to create. And if you like even mentioned to somebody a year ago, especially two years ago, but even like 6 months ago that stuff like this AI models can design, they wouldn't believe you. And here's example of what it can do on 3D. Look at this web app. Incredible 3D object animation. I can drag, do a black hole, boom, boom, boom, boom. Scroll. You can travel to zoom in. Kim K3 is really, really amazing model. Okay. So, how should you actually use Kimmy? Should you use it for everything? Should you use it with a different model as a planner as a worker? Well, cursor did this uh quick experiment of different teams of agents rebuilding SQLite, you know, the database from the 835page manual. So again, remember cost per task. This is a very clear task with a clear end objective and they tried it with one model as the planner and another model as the worker. So most expensive is Fable by far. And even though this was released before Kimmy, the point here is that look at this. Fable plus composer achieved the same result which is rebuilding SQLite for nine times almost 10 times less than just using Fable. So if you're going to use Kimmy, if you want to get the most out of it, you should use it as a planner with a faster model like Copener 2.5, Gemini 3.6 Flash, GBD 5.6 Luna, you know, maybe Kim K 2.7 code. That's actually a great combo using Kimik free as the planner and Kimik 2.7 code as the worker which is very very cheap. That's how you get the most results. You use the most powerful model, the large slow model as the planner and then a small model to execute its actions and you can get the same result for nine to 10 times cheaper. This is how you need to be efficient with your tokens. So later in the video I'll show you how to use Kimkree in your favorite harness. the secret behind Kim K3 being so good, like why is this model so efficient and so damn intelligent and my own plans on self-hosting Kim Kree within two months from now and how much that is going to cost me. Real quick, if you're watching this, please make sure to subscribe. The vast majority of you are actually not subscribed. So, if you're finding this video valuable, please go below the video and click the subscribe button. It's completely free and it helps out a lot. Okay, so let's talk about why Kim K3 is so efficient. What you need to realize is every AI model is something like an assembly line. Your problems pass through hundreds of different stations called layers, right? These are the attention layers. Each of them are adding its own piece of work. And normally in most AI models, each one of these stations aka layers dumps all of its work into one shared scratchpad where it's mixed together with everything that worked before it. So when it gets to the hundth layer or the 1,000th layer, any single insight is a drop in the ocean, right? The AI model cannot clearly hear or see its important work from earlier. Now the fix to this is attention residuals. Kimmy figured out that you can keep each layer's work separate and let every new layer use attention to pick which earlier results they should read. You know which matter and which don't. So an insight from layer 5 can actually arrive at layer 100 fully intact because it's grabbed directly, not fished from the mix of hundred of others. Now the payoff of this is that the model can train like it was given 25% more compute for less than 2% of extra latency which is incredible gains in this era where everybody is compute constrainted. Now a lot of you are probably thinking but David can I actually run Kim K3 at home and what does it take and as you can see this meme from Peter you kind of need a lot of compute okay it's not going to fit on your laptop it is a big model 2.8 8 trillion parameters. So if you want to run it fast, it's going to be very expensive. So first of all, we need to wait until July 27th when the weights are released. After that, first thing I would recommend everybody do, make sure to buy either SSD or HDD some drive and download the weights from hugging face. Okay, back them up. The US government is already preparing to ban it. There have been tweets from multiple officials all saying Kim Kry this still fable again blah blah blah. Instantly download the weights. spend like 50 bucks to just buy a hard drive with more than three four terabytes of storage and on 27th instantly download it from hugging face. That's the step one. But the real issue is that if you want to use it usably, you need kind of 50 tokens per second, right? That's kind of the minimum of usability. You need at least eight B300s. Okay, so basically you need a whole DGX B300 server with 2.3 TB of HBM memory which will set you back between 300 and $400,000. However, that's a complete distraction. We don't need 50 tokens per second. You can run this model overnight. Again, you're not going to replace 100% of your usage with self-hosted open source models. That's not the goal. The goal is to not give all your money to these closed source cloud companies. Okay, there's two main variables how we need to think about it. Two axes. You have open source versus closed source and you have self-hosted versus cloud hosted. Ideally 100% of your tokens would be self-hosted open source, right? Running on your own hardware which which you completely own running on models that are fully open source. But that's not reality. That's not possible. That's idealistic. When you're building software and when you're doing important work, you still probably need to use cloud hosted models, right? that are running in a supercomput in some data center that can run at 50 100 150 tokens per second with big models which sometimes most of the time they're closed source. So how you need to think about it is from moving 100% of your token spend being in the closed source cloud hosted corner you need to start moving it into the open-source self-hosted corner. Start by 10% 20% 30% of your monthly token spend. And for a lot of workflows, a lot of use cases, you do not need 50 tokens per second. In fact, 5 to 10 tokens per second is more than enough to cue some tasks, run them over a couple of hours, you know, prepare the work during the day and then give it to Kimik free overnight and it's going to be running on your own hardware and doing real work, sometimes better than Fable when we talk about 3D front end, legal, and much more. In fact, you can make it happen for less than $22,000 depending where you live. And here's the how that math breaks down. Okay, when I did a lot of research, a lot of deep research with JGBT, Fable, other models, deep API, the pricing was, you know, large range. CH GBT with 5.6 soul, it thought it's like $48,000. But then Fable was highly optimistic that it's possible for 12 to 13K, but this is for like very lucky liquidation pricing. Okay, dependable realistic is 16 to 22k USD before that. That's what it's going to take at minimum to be able to self-host Kim Kree running slowly at 5 to 10 tokens per second. So, if you want to actually do this, here's more specs about roughly the type of hardware you need. And if you can afford it, I would actually highly recommend it. I'm going to be planning to make more videos myself on self-hosting, how to run bigger models locally, and stuff like that. But this is my plan to invest around this, maybe a bit more up to $30,000 into hardware so I can run the best and latest open source models locally at slow speeds albeit but still I'm going to have the model no government no big tech company can take it away from me and I would highly encourage you more of you to do the same. Now, another popular benchmark in the AI space right now is browse comp. And on this one, not only does Kimik3 achieve the best result out of any model, even beating Soul and Fable aka Miffals. So, this is, you know, unrestricted Fable, even better than Fable, it achieved a higher cost at a fraction of the dollars. Look at this. The X-axis here is dollar spent, cost per task, USD. The Y-axis is the score on Browse Cop. And Kim K3 has the highest score for one of the lowest dollars spent out of any model. So this is why I believe cost per task is actually the main metric when talking about how expensive these models are. Now I already showed you how Kimmy is great at front end and 3D, but it's also amazing at chip design. Kim K3 designed its own chip to run a nanom model on its own architecture. And with a 48 hour run, it not only built and optimized it, it also verified it with open source tools. But what's most impressive is that the result actually works. The timing closes and it decodes tokens in a simulation. And again, Kim Kfrey was able to do all of this by itself. Now, let's talk about the forbidden shortcut that every AI lab takes. For years, AI research labs have known a shortcut that makes all models drastically faster, and that is linear attention. This is basically a way of processing context that keeps a fixed size memory instead of a growing one. However, a compressed memory doesn't recall precise details well, which usually means it performs well, right? So, this is why this method always used to sacrifice quality. However, Moonshot AI, the company behind Kim K3, they solved it. They used Kimmy Delta attention, which is a better way of managing that small memory and it's up to 6.3 times faster for decoding for 1 million token context. So, when a new info contradicts something that's already stored, KDA rewrites the old entry instead of piling up more on top. But they didn't fully bet only on this. They have three fast layers with KDA and then one full attention layer with perfect recall. Research findings like this coming from Moonshot AI are reasons why Kim K3 is so powerful. Even though Moonshot has way less compute than Enthropic or OpenAI, they have less researchers, less compute, and they still are able to create competitive models. So now let's get to building with Kim K3. All right. So probably the easiest way to use Kimmy is inside of Kimmy code which is their own competitor to cloud code right their own coding agent. So to install that just go to Google type in Kim code and click on the first link and then you get a single curl command which you can run in any terminal. So open your favorite terminal and just paste that in and hit enter and this will install Kim code on your machine. Now as I mentioned if you want to use Kimk free in your favorite harness I developed a skill for that. It's the Kimmy own skill. This is going to be the second link below the video. Feel free to use that. This is going to set up Kim K3 as the default in cloth code, code codex, pi, whatever your favorite harness is, you can use Kim K3 in that harness. So before I show you Kim code and how to actually use it, let me kick off a few tasks because Kim K3 can run for a while because it's a big model. So it's not the fastest, right? So I'm going to show you how to run it inside of Pi. Right now PI agent is one of my favorite harnesses. It's highly customizable. It's a very elegant, minimal, not bloated at all and it supports any model, right? So it's model agnostic. So you could just use Kimik free through open router or any other provider in the future. Again, in 5 days, the model weights are going to be open source. So we can expect an explosion of influence providers. You can ask like who are you? Let's see if it's running. Okay. Hey David, what do you need? So we're getting response. I'm Kim made by Musher AI running in your PI agent. All right. Amazing. So the first prompt I'm going to give it is something way more advanced. Not a simple front end, not a SVG design. I'm going to have it recreate Blender. Your task is to recreate Blender, the 3D modeling software, but do it as a web app. Very simple, plain English prompt. And I'm sending it off. Kim Kree with high raising effort. Let's see if it can do this. And again, this is going to run for a while. So, I'm also going to kick off a second task. And the second task is going to be something I actually need. And that is to update my hiring page so that the numbers here are live, right? Because [clears throat] u right now they're static. They're hardcoded. So, I'm going to take a screenshot and I'm going to send it. So, we're testing a 3D task because Blender is a 3D modeling software and this is going to be a front-end task. So, I'm going to say I want you to change this landing page so that first of all, redesign it completely like a worldclass front-end designer would so it looks better. And second, I want you to implement live numbers for all of the followers, right? So, start by YouTube and Twitter. Just start with these two. So it uses the actual subscriber and follower amount from my Twitter and from my YouTube. Use deep API to get the latest data and to do any scraping. So again, super plain English. I just voice prompted in 30 seconds and it's going here. The second one is asking me, do you want to blend MVP? No. Build this code testing folder and make it a clone of Blender that allows us to create 3D models and make changes to them in the browser. So, it should be a web app. Get to work and make this happen. Do not ask any more questions. All right, so I've kicked off both agents and now they're going to begin building. All right, so the build has finished. $1.3 to build a Blender clone. It's pretty crazy. And let's see. So, I'm going to say uh start the front end, open it in my Brave browser as a new tab. You can tell these models a lot more than people think. You can tell it to find specific files, to reorganize your folders, to set timers, to run caffeinates so your computer doesn't go to sleep, to change settings on your machine, to debug your network when your internet is terrible at your office. You can do a lot with these AI agents. Chances are you're probably underestimating how much these agents can do. So, you know, I could al tap and open a new browser tab, but what is this? You know, are we in the 1900s? I'm not going to be doing some manual labor here. I'm just going to have it open Blender for me. and uh open u the Blender clone in a new browser tab. So, it did that. Let's see what the controls are. Uh uh uh. Okay. So, we have a cube. This looks like Blender. Reminds me of my good old gamedev days. So, this is the scale. Let's see. Okay. We can make the scale larger. And it works. Uh what are G transform? Okay. G Rotate. Okay. S scale. Nice. This is nice. Object mode, edit mode. Yeah, I mean, I forgot the Blender controls, but this is solid. This is very solid. It has the basics. Obviously, it doesn't have all the settings, but again, with enough time, you could probably get all the settings to work. Like really, it's it's all about like being able to define it and having the right software design and architectural skills. Rather than remembering syntax, you need to be a better architect and know how to guide these models. like you really need the skill set of a developer who would manage like 10 junior developers. Right now it's like with Kim K3 it's more like 10 senior developers but still you need to be more like a development manager more like a CTO of a company being able to divide work make sure people are doing the right thing know how to check the quality quickly and efficiently with limited time. Uh so yeah this is the 3D stuff but let's look at the website. Did it push? Uhuh. Uh yes start next death. So let's look at it. We can I yolo to pro probably. So right now there's only one provider and that is the Kimi platform. So if we go to Kim K3, you can check that on open router. There's only the moonshot AI provider, right? But this will completely change once we have the weights in 5 days. We're going to have way more providers. Again, some of them are going to have higher throughput. Some of them are going to have a better pricing or lower latency. You can see that the moonshot has amazing uptime better than enthropic and open AI. But this is the main benefit of open models. Anybody can run the inference any anybody can download the weights and then anybody can develop better ways to run the model. Anybody can fine-tune the model. It really is the way to make the future better is open the research and open the technology. All right. So Kim is starting the server. So it's on localhost 3001. Right. The server has finally started. So let's switch to Brave browser. And here it is. So, here's the old one. Incorrect numbers. And the new one. Oh, look at this. Nice animations. Clean, clean, clean. I mean, some of this has been already there, but the design is definitely better. Kimmy has that taste. Oh, very nice animation there. Life amounts. I think it's redundant here. So maybe I'll have it remove this. Uh uh uh. So again, feedback super simple. Attached image. Remove the top ones. This is redundant. Only keep the main big numbers in the same strip where it's uh next to the vector users as well. Now, while it's making that one fix, let me show you how to set up Kimmy code. So again, we ran this command. And now we should be able to just type in Kimmy. There it is. You need to open a new terminal. It wasn't in this in this one. So we have Kimmy uh CS installations migrate data to Kimmy code. Yes. Migrate now migrate chat sessions. Okay let's do that. Enter. And here it is. Kim code is right here. Uh uh let's to access my folders. Oh my bad. I posted it in the wrong chat so it was confused here. So I'm going to close the blender so we have a bit more space. And let's focus on Kim code right here. So Kim code. Let's do model Kim for coding. Okay. Who are you? And we need to first login. So do SL login and we have multiple options. Kimmy code or Kimmy platform if you want to use API key. So if you already have a subscription which they had so much demand that they had to pause it. So when you go to kimmy.com/code, you can still see that uh subscriptions are here. So if you want to get the most value, get a subscription for sure. This is way more value than the COX one or the Enthropic one. But in fact, this might not be available forever because they literally had to pause the subscriptions few days ago because of the demand, right? Because the Kim K3 was such a hit. They don't have infinite compute. They have far less compute than American companies. So, they had to pause the subscriptions. So, I would get this while it's still available. But, if you want to use the API platform, and again, if you use the first link, you're going to get 15% of extra credits. We can also do that. So again inside of your terminal select API key and select the third option uh the platform.kkey.ai enter. And we need to paste an API key. So go back to your platform go to API keys and click on create API key. I'm going to name it K3 testing live. Select a project. Do okay. And just copy it. Again do not share your API keys with anybody. Treat them as passwords. Okay. I'm going to delete mine before uploading this video. So we have the models. So obviously select Kim K3. It's the best. And you can uh switch the effort to slasheffort. So low, high, max can keep max. And let's send a test prompt to see if this works. And we are seeing tokens. So I do think this works. I'm keeping CLI coding agent running on your Mac. Amazing. So actually I'm going to escape it. Ctrl + C. Let's launch it in a project. CD code testing. Boom. Boom. Just type in Kimmy. Here it is. Kim K free thinking max. And as you can see, Kim Code also has scheduled tasks. So you can do like remind me at 5:00 p.m. stuff like that. Um to kind of automate it a bit more. So to test this, instead of showing you more coding tasks, we're going to do something unconventional. And I'm going to show you the OpenAI lawsuit, right? Apple suing OpenAI against uh stealing of the hardware because again Kimi is the best model in the world at legal tasks and by the way that's not by little the jump is massive right this is the Harvey lab from the AI tool Harvey basically even though this uh image is a low quality fable scores 14.2%. And it's the second best model in the world. Grock 4.5 surprisingly good at 13.3%. But Kim K3 26.7%. Basically double the performance of Fable at this uh legal benchmark. So anything where you need to review a contract or you know do some legal research use Kimmy K3. So what I have is like find the PDF in this order about the Apple X Open AI lawsuit. First I'm going to tell it to just find the PDF and then we can ask it many different things, right? such as the strategy it would take in the case of OpenAI or in the case of Apple. So I'm going to say read it in full one page at a time and tell me what side has the stronger case and if you had to bet what is the likelihood that Apple wins this lawsuit. Answer in short in plain English. I've been ending nearly all of my promps by the way with this answer in short in plain English. Both with GBD 5.6 soul both with Fable 5 and with Kim K3 as well. Uh okay. So we need to approve for the session. We're not running in the YOLO mode. So we need to do approvals. Can actually ask use deep API to browse the web to figure out how to run Kimi code in YOLO mode. By the way, if you don't know what deep API is, this is my next big project I'm working on. I've been keeping this in stealth for the past three or four weeks, but we're finally ready to review more people and open it to the next stage. It's not going to be a full public release. So if you want to have access, we're going to review it and we're going to accept the first, you know, 10 20 people in. So go to deep aapi.co if you're interested in having a single API key that can do everything. Anyways, uh let's go back here approve for the session. I I might want to restart this in yellow mode so I don't have to approve everything. But we have two the sessions of Kim free running. One of them is researching the web. The other one is uh already scanning the PDF. All right, it read all all foreign pages. Apple has a stronger case clearly. I mean this is uh you know written from Apple's side. So that is expected that is worded in Apple's side. But let's see my bet. If the messages are real and logs are real, Apple is very likely to wig against Leo. So these are some people I guess who switched to OpenAI who worked at Apple and they went to OpenAI against OpenAI corporate proving institutional liability is harder. I put Apple over meaningful win as roughly 70 to 75%. Okay. Most likely outcome settlement or jurisdiction against the individuals and ops practices, not a full trial. Yeah, as it says, this is only Apple's side of the story. So, obviously, it's going to be biased towards that. But again, if you're moving to a different city, signing a new apartment, have Kim Kree review it to make sure the law landlord doesn't screw you. If you're signing an office contract, have Kim Kree review it. All these things, all these legal tasks, Kim Kree is currently the best model in the world. and not by a small margin, by a very, very substantial margin. So, this has been Kim Kree. Again, a huge thank you to the Kimmy team for sponsoring this video. And if you want to try Kimmy yourself, either inside of Kim code or building on top of it, you know, building your own software on top of it, application startups, make sure to use the Kimmy API, the first link below the video will sign you up. And if you're a new user of the API, you're going to get additional 15% of credits extra completely for free. And if you want to use Kimik free in any agentic harness, whether it's codex, cloth code, pi, or anything else, click the second link below the video and get this free skill which will set that up for you. Again, it's completely free. So go ahead and grab it