Kimi K3 Beats Claude on Coding — But There's a Catch

[00:00:00]China just dropped an open-source AI model so big it's making Wall Street Journal nervous again.

[00:00:07]This is not just hype. This is the same panic moment we saw with the DeepSeek, except this time the model is nearly three x bigger, and it's beating Claude and GPT on part of the coding leaderboard. Today, I'm going to break down Kimi K3, real number, real reaction, and whether it actually matters for you or not. Kimi K3 is the open-weight model from the Moon Shot AI, a Chinese startup backed by Alibaba, released in mid-July 2026.

[00:00:40]It's a 2.8 trillion parameters model.

[00:00:44]That makes the largest open-source AI model ever released, bigger than DeepSeek version 4 1.6 trillion model.

[00:00:55]It's not a dense model, though. It's a mixture of expert. Only 16 of 896 expert activate per token. So, the actual active compute is closer to 50 billion parameter. That's a trick that makes the model this size actually usable.

[00:01:16]It also ship with a 1 million token context window and native visual understanding. That's incredible. Here's the part which actually blew up online.

[00:01:28]K3 didn't just enter the leaderboard, it topped it. Number one on the front-end code arena, ahead of Claude. It won six out of seven front-end subcategories, only losing the game category. But, this is the part clickbait headline skip. On the general text arena, it only ranked around sixth to ninth.

[00:01:53]So, it's not the smartest as the model ever overall. It sit between the Claude Fable 5 and GPT 5.6 Soul.

[00:02:04]But, ahead of Claude Opus 4.8.

[00:02:08]Ranking fourth overall on the Artificial Analysis Intelligence Index. The first open weight model ever cracked the top four. Let's go rapid fire through the number Moon shot actually published.

[00:02:22]88.3% on Terminal Bench 2.1.

[00:02:27]93.5 on the GDP at Diamond, the part open weight result at launch, and a state-of-the-art 91.2 on Browse Com.

[00:02:38]That GPQA score matters if you want to have a self-hosted model for hard reasoning task. Now, here's the catch nobody pulls in the thumbnail.

[00:02:49]K3's accuracy climbed from 33% to 46% compared to its preceder.

[00:02:57]But, its hallucination rate climbing to from 39% to 51%.

[00:03:05]K3's got smarter and more confidently wrong at the same time.

[00:03:11]>> [laughter] >> It answered more questions correctly, but also make up wrong answers with confidence.

[00:03:18]If your workflow depends on the model knowing when it does know something, there is a problem. Now, things are getting interesting. The release coincident with the Xi Jinping's speech at the world's AI conference in Shanghai, and semiconductor stock sold off on fear that cheap open Chinese models undercut the whole you need Nvidia chip and close model. Now, look at this exposed, where David Sacks, the former White House AI crack is reacting.

[00:03:54]He's basically saying US regulations is what actually losing the race, not the tech gap.

[00:04:01]But not everyone agree it's a big deal.

[00:04:05]Now look at this other post. Patrick Moorhead, a chief analysis to calling the reaction on over reacting comparing it directly to the deep fake panic. So you've got one communication saying you're losing another saying calm down this happens every few months.

[00:04:25]One more post worth showing this one's just fun.

[00:04:30]Now look at this exposed from developer who posted a photo of Moonshots actual office two days before the launch. A company worth tens of billions building a model beating Claude on the leaderboard working out of what looks like a normal office.

[00:04:49]The contrast is the whole story of Chinese AI labs right now.

[00:04:54]Real talk. Should you use Kimi K3?

[00:04:59]Comparing the pricing roughly $3 input and $15 output milli per million token.

[00:05:07]That's about triple what the previous version cost but still cheaper than Opus 4.8 or Fable.

[00:05:16]If you are doing front end or UI generation, it's currently the strongest open open model on the specific leaderboard.

[00:05:27]If you need something that won't confidently make up things, the hallucination says be careful.

[00:05:34]So bigger open source model ever. Number one on the major coding leaderboard and it's rattling US tech stocks. Is it actually better than Claude or Jet Chat GPT? Depends on what you are using it for. And now you have got the real number instead of the headlines. If you want me to build something with Kimi K3, let me know in the comment section and for more amazing AI informations, stay tuned, hit subscribe and see you in the next video. Till then, goodbye.