Open-weight AI just hit 2.8 trillion parameters…

[00:00:00]Last week, Chinese AI lab Moonshot dropped Kimi K3, a massive open-source monster that instantly parameter mogged every other open model in existence. And not just that, it has OpenAI and Anthropic terrified because its trust-me-bro benchmark performance is on par with, and in some cases beating, Claude Fable and GPT-5.6 Soul. And that's crazy because just a few weeks ago, the government was telling you these models were too dangerous for the common man to use. It only took the Chinese a few weeks to match their performance and then give away all the weights for free. But big AI is not too happy about this, and there are already calls to ban Chinese models completely in the United States. In today's video, we'll take a look at the massive breakthrough that is Kimi K3 and find out what it means for the future of artificial intelligence. It is July 22nd, [music] 2026, and you're watching The Code Report. Like Claude Fable and GPT-5.6 Soul, Kimi K3 is a native multi-model mixture of experts model with a 1 million token context window and a staggering 2.8 trillion parameters, and is optimized for jobs like long-horizon reasoning and, of course, coding. One interesting characteristic of K3 as a mixture of experts model is that it has 896 total experts, of which exactly 16 activate per token, which means it works just like a big corporation where 16 good programmers do all the work while 880 other managers sit there and do nothing. But apparently, this works out well for Kimi because it makes scaling about 2.5 times more efficient than K2.

[00:01:23]Despite these gains in efficiency, Kimi was so popular upon release that their GPUs ran out of juice and they had to start turning away paying customers. And as of making this video, all the paid plans are currently sold out. That's unfortunate, but in theory, you could self-host the model because the weights are open. The weights are expected to be released on July 27th, but there's no chance in hell you'll be able to run it on your little gaming GPU. To run a monster like this, you'll need a massive array of data center caliber GPUs. But in theory, if you could do that, you would have unlimited access to a Fable Soul caliber model, and that would be amazing because the benchmark situation with K3 is pretty wild. K3 is ranked number one on front end code Arena at a 1,679 Elo, which puts it ahead of Fable 5 and GPT-5.6 Soul. In addition, it lands in the top three on the artificial analysis intelligence index. And if we look at every other coding benchmark, it's at least very competitive with the other frontier models. But you should never trust the trust me bro benchmarks because many of the K3 numbers were produced with Moon Shot's own Kimiko harness while competitors ran in different harnesses. That could make Kimiko look slightly better at coding, but to their credit, Moon Shot admits that K3 still trails Fable and GPT-5.6 Soul overall, especially on benchmarks like Humanity's Last Exam where it's down by about 10 points. On top of that, artificial analysis measured a 51% hallucination rate, which is definitely not a good thing, especially when it comes to coding. In addition, it also tends to spit out way more tokens than it needs to, which could ultimately end up costing you more money despite the model itself being cheaper. When it comes to things like UI design and data visualization, it's extremely impressive for an open model, but in my opinion, it's still one step behind Fable and GPT Soul. But one of the most interesting things about this release is the geopolitics surrounding it. Recently at the World AI conference, China's Communist Party became the loudest advocate for free and open artificial intelligence. Meanwhile, in the land of the free, Silicon Valley wants to regulate and gate keep it by pushing the fear narrative that it's about to take all of our jobs. In Washington, they're reportedly considering entity listing Chinese AI labs and OpenAI's Dean Ball argued that open weights are inherently decelerationist.

[00:03:28]>> What?

[00:03:30]Bro, what are you talking about, man?

[00:03:32]>> And coincidentally, that's very similar to the argument Steve Ballmer used to make in the '90s about Linux when he said Linux is communism. And also coincidentally, both of these guys have balls in their names. Frontier labs don't like open models simply because they divert the flow of money from them to someone else. As of today, the odds the US government bans Chinese models is only sitting at 29% on Poly Market, but that could change quickly if they determine one of these models is responsible for some kind of cyber attack. But, the best thing about K3 is that it pushes the arms race forward.

[00:04:01]Alibaba also just released Qwen 3.8, which itself has 2.4 trillion parameters and open weights. And I think this model might finally be the one that gets the UI right for Horse Tender. But, before you let AI slop out your UI, you need to check out mobbin.com, the sponsor of today's video. I've been using Mobbin for over 5 years now because it provides highly detailed breakdowns of every screen in thousands of popular web and mobile apps. And they just launched an MCP server, which connects your AI agent to over 600,000 screens and user flows from apps in every category. This gives your agent real-world references, so it can design high-quality UIs for your specific use case, instead of just spitting out generic purple gradient vibes law. It also lets you do deep UI research, like you can feed it your prototype and your agent will use Mobbin's library to provide a ranked list of other apps that do it better.

[00:04:50]You can also ask it how your competitors handle things like onboarding and paywalls, and it'll show you every screen in their full user flow. And so, if you're tired of your UIs looking like the same as everyone else, I'd highly recommend trying out Mobbin at the link below. This has been the Code Report.

[00:05:04]Thanks for watching, and I will see you in the next one.