Kimi K3: The Free AI That Just Beat Claude at Coding (Ranked )

[00:00:00]Kimi K3 just landed at number one on the front-end code arena above Claude Fable 5 above GPT above everyone else and the whole thing is free and open source, but you probably don't know why an open source model like this changes everything for the business you're trying to build. So here are five things about Kimi K3 that genuinely shocked me with the receipts. Let's go. The first thing, it didn't just compete with the closed models at one specific thing, it beat every single one of them. So here's the screenshot that made me stop scrolling. The front-end code arena.

[00:00:34]This is where models go head-to-head on real front-end coding and sitting at number one above everybody is Kimi K3 1679.

[00:00:43]And this isn't some easy multiple choice quiz. The front-end code Arisa pits models against each other on real front-end builds, actual interfaces, actual code that has to work and people vote on which one did it better. It's about as close to can this thing actually do the job as a benchmark gets.

[00:01:02]Look who's underneath it. Number two, Claude Fable 5. Number three, GPT 5.6 Soul. Then GLM, then Grok, then Claude Opus 4.8. Some of the most expensive, most locked-down models on the planet and an open source model just put its name on top of all of them. And here's the part that's easy to miss. Scroll down to number 18. That's Kimi K2.6, their own last model, sitting at 1515.

[00:01:28]So in one release they didn't inch forward, they jumped from 18th place to first. That's 164 points in a single generation. Now, I want to be straight with you because I'm not here to hype something that isn't real. This is the front-end code arena specifically. On their own site, Moonshot publishes a full benchmark table and it's honest. K3 wins program bench at 77.8.

[00:01:53]It wins the SWE marathon at 42. But on some of the broader tests, Claude 3 5 and GPT-5.6 still edge ahead. So, the fair read isn't Kimi is now the best model in the world. The fair read is an open model you can download for free is now trading punches in the same ring as the giants. That's the story. But, topping a leaderboard isn't even the crazy part. The crazy part is what Moonshot calls this thing in their own words, and it's something no model this good has ever actually been. Right here on their announcement, they describe Kimi K3 as, and I'm quoting them, "The largest open weight AI system ever built." 2.8 trillion parameters, a 1 million token context window. Let me translate what open weight actually means for you, because this is the whole ballgame. When you use Claude or GPT, you're renting. The model lives on their servers, behind their door, on their terms. If they change the price, you eat it. If they change the rules, you adapt.

[00:02:55]You never actually hold the thing. Open weight is the opposite. Moonshot is saying, "Here are the actual weights, the brain of the model, and by July 27th, you can download it, keep it, run it. Nobody can take it back from you or quietly nerf it on a Tuesday." And this isn't a little research toy, it's the biggest one anyone's ever released. 2.8 trillion parameters of frontier-class capability, and the plan is to just give it away. Think about how strange that is. [music] The most guarded thing at OpenAI and Anthropic is the model itself. Moonshot's move is to hand it to everybody. So, a model this powerful is about to be free and open, which means the only question that actually matters for you is how do you get your hands on it, and what does it cost? And this is where it stops being a headline and starts being useful to you. The pricing.

[00:03:49]$3 per million input tokens, 15 per million output. That's Claude Sonnet territory for output that's competing with the flagship models. Now, could you run the full 2.8 trillion parameter monster on your laptop? No, that's a serious machine. But, here's the thing, you don't have to, and I don't [music] want you thinking this is out of reach because it absolutely isn't. You've got two easy paths. One, just use it on Kimi's own website or their app or their playground. It's live right now. You log in and you're using it. Or two, and this is what I actually did. I grabbed their API key. I dropped it into my own agent operating system, Agent Flow, and now Kimi K3 just shows up in a drop-down menu right next to every other model I use. So, when I fire up my agents, I can point them at Kimi with one click.

[00:04:39]That's the unlock. One API key and this brand new frontier model is running across all my agents. No migration, no rebuild. I select it from a menu and go.

[00:04:49]Every time a new model drops, that's the move. Plug the key in, pick it from the list, done. So, it's cheap and it's genuinely in your hands, which raises the real head-scratcher. How is a company most people have never heard of doing what OpenAI spends billions on? It comes down to one clever trick cuz normally, the way you make a model better is brute force. More GPUs, more data, more money, the OpenAI playbook.

[00:05:16]Moonshot did something different and they explain it right on the page. Their words, "Kimi K3 is built on a hybrid linear attention mechanism they call Kimi Delta attention." And the payoff they claim, quote, "Up to 6.3 times faster decoding in million token context." Let me put that in plain English. Attention is the expensive part of an AI. It's how the model holds everything in its head at once, and it gets brutally slow as the context gets huge. Kimi's trick makes that part dramatically cheaper, so they can handle a million tokens without the cost exploding. And a million tokens is a lot. That's your entire code base dropped in at once. That's a whole book or months of your businesses documents or every transcript you've ever recorded all in the model's head at the same time, and it can still move fast through it. For an agent that has to actually understand your business before it does anything useful, that's massive. And there's a second piece they're proud of called attention residuals. They're claiming about 25% higher training efficiency while adding less than 2% in extra compute. 25% more for basically free. Stack that up and they say the whole system converts raw compute into usable intelligence far more effectively. Two and a half times better scaling than their last model. That right there is the answer to how a small team competes with giants. They didn't outspend anyone. They out-engineered them. They found a smarter way to turn electricity into intelligence, and that trick didn't come out of some thousand-person lab, which brings me to the part I still can't wrap my head around. Who actually built this?

[00:07:01]Moonshot AI was founded in March of 2023. That's it. 3 years ago by three schoolmates from Tsinghua University.

[00:07:09]Three people. Three years from a standing start to a model sitting on top of the front-end code arena above Claude and GPT. Investors clearly see it, too.

[00:07:20]They're reportedly racing at around a $31 billion valuation, >> [music] >> and don't gloss over the open-source part because it's a genuine shift. The frontier used to belong entirely to a handful of American labs, and it was all closed. Now, the model at the top of this board is one you'll be able to download for free that pulls the whole floor up. Every builder, every small team, every operator suddenly has frontier tools without a frontier budget. That's a very different world than the one we were in even a year ago.

[00:07:52]And I think there's a lesson buried in here that's bigger than the model. They didn't win because they had the most, they won because they had constraints and constraints forced them to be clever. They couldn't outspend OpenAI, so they had to outthink them and it worked. Three people, three years. So here's the real question. What does a tiny open team beating the giants actually mean for you and the business you're running? Because that's really who I make this stuff for, not just creators, anybody who wants AI agents doing real work inside their business.

[00:08:28]Maybe you're a creator, maybe you run something way more hands-on, doesn't matter. If you want agents doing the heavy lifting for you, this release matters and here's why. Up until now, the the best AI was a tax. Every agent you ran, every task it did, you paid a toll to a company you don't control on a model you can't keep. That's a scary thing to build a business on top of.

[00:08:51]Kimmy K3 cracks that open, near frontier capability at budget pricing on weights you can actually own. The engine behind your agents just got dramatically cheaper and a lot more yours. And think about what you can point that engine at.

[00:09:06]You could run agents that pump out content and pull in leads. You could package what you know into a community and let agents deliver it. You could build an app that sells agent workflows to other people. You could run a whole agency selling these skills, doing it for your customers with the agents doing the work in the background. I don't care which one you pick. The point is the same. The thing that runs it, the actual intelligence, just stopped being a locked, expensive black box and started being cheap, open, and yours. When a three-person team can top the leaderboard, the story you've been telling yourself about needing a big budget or a big team to build with AI, that's over. If you're building this stuff and you want people figuring it out alongside you, come hang out in the community links below. And if you want to see the exact setup I used to run every one of these models, watch this one next.