China's Qwen 3.8 Max DESTROYS Kimi K3?

**Julian Goldie** (0:00)
We got another big launch today. Qwen 3.8, second only to Fable 5, according to Qwen's announcement here. It's open source, going live soon. It's not available to CodeWeave unless you're in mainland China. I did check out the international website, as you can see, but Qwen 3.8 is not available on there just yet. However, what I will say is this is a big one, a big one, my friends, because we have a 2.4 trillion parameter model, which is absolutely massive. The model is continuously evolving. They believe, according to Qwen, and just test this stuff out for yourself, don't believe everything you read, as always, but they believe it's one of the most powerful models available today, compatible to leading frontier AI models. Second only to Fable 5
Now that is a huge statement because number one, it leapfrogs Gpt 5.6 sole, and number two, this has already, according to them, beaten Kimi K3. This was an announcement just made a few hours ago, so literally about one hour and 10 minutes ago, and pretty amazing stuff to see this. Now, Chinese AI is controversial. I mean, it is overtaking very quickly stuff like, for example, Gpt 5.6, Fable 5 is only available on 50% of the usage plan of Claude, which means you don't really properly have access to it, and these open-source models are free. So for example, I mean, you're not gonna host them yourself. They're too big to host yourself. You need a ridiculous amount of DGX Sparks connected together. But if you look at probably the leading model in the world right now, which is Kimi K3, I should be on my tests.
A lot of models, for example, like Fable 5 and Gpt 5.6. So on the benchmarks, these models win. But on the actual tests that I do personally on GoldieBench, which you can check out on goldiebench.com, Kimi is crushing them. So if this comes out and it's second only to Qwen 3.8, we are looking at a very strange world where two Chinese models have leapfrogged ChatGpt in the space of a week. Bear in mind, there doesn't seem to be like any preview, you know, preview sort of system where, for example, if you look at Gpt 5.6, you couldn't access it for two weeks. It was only available to like cyber security firms. If you look at, for example, Fable 5, comes out for three days, gets shut down, comes out again, never really feels quite the same, and you can only access it on a limited basis, a very limited basis. And we don't know if it's gonna be on the core plan forever. However, you look at Chinese models, open source models, like for example, Qwen 3.8, like for example, Kimi, they're available for free. They're going open sourcing. Access them inside the chat at kimi.com for free. And also they perform very, very well. In fact, to the point where, for example, if you look at Kimi K3, it's actually outperforming on coding benchmarks, other models like, for example, Fable 5 It's also top of the front end coding benchmarks. You can see it is crushing Fable 5 and Gpt 5.6. It's just a weird week. Like this is everything is changing. Like the tables just turn every day. And we are looking at a situation here where the Chinese models are just getting ahead of the US models. And there's going to be more and more sort of restrictions on open source models potentially in the future. This actually brings me on to something that was looking at from Dean W. Ball, head of strategic futures at OpenAI. And he did like quite a really interesting post here.
Obviously like Gpt 5.6 or the OpenAI team, they are worried about Chinese models. You got Fable 5, who I'm sure are worried about Chinese models as well. There's also been rumors of like distillation as well. However, what Dean was saying on the latest update from Kimi K3 this week is that it's a very good model. However, he doesn't think that its performance is explained by distillation. So distillation is kind of like where you test a model thousands of times and then you understand the patterns. And then you train your own AI model to think in that way. And the very simplified explanation, that's how it works, right? I'm sure it's much more complicated in reality. But basically what they're saying now is that Kimi K3 seems pretty much on par with the best public models. This is the head of Futures at OpenAI, so it's not like some random geeks are on Twitter saying this. And so he also said it's very token hungry and not that cheap to run. But actually, if you just use it on the Coding Plan, it is cheap to run because then do you have your set limit, you're not paying per API. Now also, he's also said that he's personally surprised the Chinese state continues to allow the open sourcing of models. This could give in the potential risks. However, like if you look at the OpenAI team, they number one have a lack of inference, right? Because for example, like there's all these export controls on chips going to China, which means like if they open source it, then people just host it themselves. Or you get for example, Olamar and companies like that, hosting Kimi and then giving it away to people, which means like these open source models can be distributed more without China actually having the chips to deploy them and to make it publicly available as much. Then you also got an interesting situation here, where basically he said the one probable outcome of an open weight model dominant world, aka, and this was before Qwen 3.8 was announced, but now after Qwen 3.8 has been announced today, this goes even more wild, if that makes sense, because he said basically one probable outcome of an open weight model dominant world is cool AI communism, which is an interesting one where basically AI is public good, which is ultimately provided by the state as a kind of digital public infrastructure.