Somewhere in Beijing, a startup you've probably never heard of just dropped an AI model so big it made semiconductor traders on three continents spill their coffee. Moonshot AI didn't just ship an update — it shipped a shockwave, and global markets are still doing the math.

2.8 Trillion Parameters Walk Into a Benchmark

Moonshot AI released Kimi K3 on July 16, a mixture-of-experts model built on roughly 2.8 trillion parameters — now the largest open-source AI model on the planet, about 75% bigger than DeepSeek's V4 Pro. Despite the size, it only activates 16 of its 896 total experts per task, so it's less "brute force" and more "extremely well-organized committee."

The model grabbed the No. 1 spot on Arena.AI's Frontend Code Arena, edging out Claude Fable 5 and GPT-5.6 Sol on head-to-head coding comparisons, and topped four of eight real-world automation benchmarks. Pricing lands at $3 per million input tokens and $15 per million output — squarely mid-tier by Western standards. Full model weights are due July 27.

Wall Street Did Not Take This Well

Within hours, Bloomberg's Asian semiconductor index slid more than 6%, Taiwan's benchmark dropped over 6%, Japan's market closed down 4%, and Nasdaq 100 futures fell 2% before the US even opened. Chinese AI rivals got caught in the crossfire too — Zhipu cratered 28% and MiniMax fell nearly 16%.

The real story isn't that Kimi K3 is the best model out there — it isn't, and it still trails Fable 5 and GPT-5.6 Sol on overall performance. It's that a startup can now ship frontier-adjacent capability at commodity prices, which is exactly the sentence that makes anyone holding chip stocks reach for the antacids.

Moral of the story: in the AI arms race, "open source" and "state-of-the-art" just moved in together, and the rent is suspiciously cheap.

Source: CNBC