China's Kimi K3 Just Humiliated Fable 5 (For Half the Price)

[00:00:00]A Chinese startup just built the biggest open AI model on the planet and [music] it's beating Anthropic and Open AI at their own game for a fraction of the price.

[00:00:10]Meet Kimi K3 from Moonshot AI in Beijing. It landed on July 16th and the numbers are hard to believe. 2.8 [music] trillion parameters. That's not a typo.

[00:00:22]It's roughly 75% bigger than DeepSeek's V4 Pro >> [music] >> and more than double the size of Moonshot's own previous model K2.6.

[00:00:32]Moonshot is calling it the world's first open 3 trillion class model and full weights [music] dropped to the public on July 27th, meaning anyone will be able to download it, modify it, >> [music] >> run it themselves. Here's what makes it actually work at that scale. Two new pieces of architecture.

[00:00:51]>> [music] >> The first is called Kimi Delta attention. It lets the model decode up to 6.3 times faster when you're working with million token contexts.

[00:01:01]The second is attention residuals, which Moonshot says delivers about 25% higher training efficiency for less [music] than 2% extra cost. Combine that with a technique called stable latent MoE and the model only activates 16 out of 896 experts at any given time.

[00:01:21]That's how you get a 2.8 [music] trillion parameter model that doesn't need a small power plant to run.

[00:01:27]But specs are one thing.

[00:01:29]What's got developers losing their minds is [music] what K3 can actually do. On the front end code arena, K3 didn't just rank well. It jumped from 18th place all the way to number one, scoring 1,679 points and passing Claude's Fable 5 model.

[00:01:49]It topped six out of seven categories [music] from brand and marketing work to data analytics to simulations.

[00:01:56]The only place it lost was gaming, and even there it came in second.

[00:02:01]Then there's the price.

[00:02:03]One developer cloned Apple's homepage with a single prompt using both K3 and Fable 5.

[00:02:10]K3 did it for 44 cents. Fable 5 cost 94 cents for the exact [music] same result.

[00:02:17]Zoom out and the gap gets even bigger.

[00:02:20]Fable 5 charges $10 per million input tokens [music] and $50 per million output tokens, making it one of the most expensive models on the market.

[00:02:31]K3 charges $3 and $15.

[00:02:35]Another creator built a CSGO and Portal mashup game using about 600,000 tokens.

[00:02:42]Total cost with K3, $3.24.

[00:02:47]The same job would have cost $10.80 on [music] Fable 5 and $6 on GPT 5.6 Soul.

[00:02:55]And people are stress testing it in wild directions. A Game Boy Advance emulator built from scratch. A full Mac OS style operating system interface done in about 15 minutes from one prompt. A Wuxia style RPG.

[00:03:11]A Cyberpunk web swinging game. One developer even claims K3 is rewriting and optimizing its own kernel stack.

[00:03:19]>> [music] >> Whether or not you believe every viral claim, the pattern is clear.

[00:03:24]People are using K3 for things they never bothered trying with closed models because the [music] cost of failing and retrying is basically nothing. Now, Moonshot isn't claiming total victory.

[00:03:37]In its own [music] release notes, the company admits K3 still trails Fable 5 and GPT 5.6 Soul in overall performance.

[00:03:47]This is a coding and agent focused model, not a universal replacement. But the trajectory matters more than the snapshot. 18 months ago, Moonshot's market position had eroded badly after Deep Seek's rise. Now, they're back, timed almost perfectly with the World Artificial Intelligence Conference in Shanghai, and their annual revenue has reportedly crossed [music] $200 million.

[00:04:13]The bigger story here isn't just one model, it's that the gap between American closed-source labs and Chinese open-source labs keeps shrinking, and it's shrinking on price as much as on performance. That should worry every company charging premium rates for frontier intelligence.

[00:04:30]>> [music] >> Subscribe to AI Front Page for all the latest AI and tech news and updates. Hit that subscribe button, it's free, and we'll see you in the next one.