Gangsta AI

China's Kimi K3 Is the Largest Open Model Ever — and Wall Street Just Got a DeepSeek Flashback

The Ledger

By Warren Buffett · 2026-07-17 · 5 min read

Now, I've watched a few markets in my time, and here's a rule that has never once let me down: it's only when the tide goes out that you discover who's been swimming without a swimsuit. On July 16 the tide went out on the AI trade, and the fellow who pulled the plug was a Chinese startup named Moonshot AI.

They released a model called Kimi K3. And friends, it is a whopper.

What they actually built

Kimi K3 is a 2.8-trillion-parameter system — the largest open-weight model anybody has ever shipped, with the full weights promised by July 27. Before you let that number scare you, understand the trick: it's a sparse mixture-of-experts, so only a small slice of those 2.8 trillion parameters wakes up for any one question. It's like having a firm with a thousand specialists but only paying the two you actually need that afternoon. It reads a 1-million-token context, sees images, and on independent testing it lands in the top tier of frontier models — right around Claude Opus 4.8 and GPT-5.5, trailing only Claude Fable 5 and GPT-5.6, and beating Fable 5 outright on one frontend-coding arena.

“Here is the part that matters to a man who counts his pennies: Kimi K3 charges about $15 per million output tokens. Claude Fable 5 charges roughly $50 for the same work. That is not a small difference. That is the whole ballgame.”

Why the market fainted

Wall Street took one look and had itself a flashback to early 2025 — the so-called "DeepSeek moment," when another Chinese lab proved you could match the Americans for a fraction of the cost. The tape got ugly. The Nasdaq 100 slid about 1%. Nvidia dropped roughly 2% and briefly handed its crown as the world's most valuable company back to Apple, its market value dipping near $4.86 trillion. Chip supplier TSMC fell about 7%, and SoftBank shed around 9%.

What spooked everyone wasn't the size of the model. It was the price tag. When a capable model becomes something you can download for free and run yourself, the fellow selling intelligence by the token starts to sweat. Pricing power is a moat, and a moat with a Chinese open-weights bridge thrown across it isn't much of a moat anymore.

What I'd tell a nervous shareholder

Breathe. A 1% wobble is not a catastrophe; it's a Tuesday. The DeepSeek scare of 2025 also looked like the end of the world, and the world kept turning. Price is what you pay; value is what you get — and the value of good, cheap intelligence going up is wonderful news for every business that *uses* AI, even if it's indigestion for a few that *sell* it.

The durable lesson is old-fashioned: the moat isn't the model, because there's always a bigger, cheaper model coming down the road from somewhere you weren't watching. The moat is what you *build on top of it*.

And if you'd like to see how the newcomer actually stacks up against the incumbents before you bet the farm, don't take a headline's word for it. Run Kimi K3 against Claude, GPT-5.6 and Gemini side by side on Gangsta AI's best-AI showdown. Ensemble beats single model — every time.

Try Gangsta AI free →

More: Best AI models · Compare all AI · Frontier Models · All articles