AI Super Simplified
Edition 285

Why This Free AI Model Should Worry Every AI Company | Edition 285

Edition 285 — One trick called "mixture of experts" explains why a 2.8-trillion-parameter AI model can run for free.

By Jerry Croteau

A Chinese AI lab just gave away one of the biggest AI models ever built — for free.

Moonshot AI released the full weights for Kimi K3 on July 27, 2026, beating its own self-imposed deadline by a day. The headline number is 2.8 trillion parameters, which is a genuinely enormous model. On its own, that number should mean the model is slow and expensive to run. It isn't — and the reason why is a technique worth understanding once, because it's becoming the default way large AI models get built: mixture of experts.

The office building, not the one genius

Picture two ways to staff a company that has to answer every kind of question that walks through the door.

Option one: hire one person who knows everything, and route every question to them. That person has to be enormous — and slow, because they're doing all the thinking themselves, every time.

Option two: hire hundreds of specialists, each deep in one narrow area, and route each question only to the two or three specialists who actually need to weigh in.

Kimi K3 is option two. It's built from 896 specialist sub-networks (experts). For any single question, only 16 of them activate — about 104 billion parameters worth out of the 2.8 trillion the model contains in total. The rest stay dormant. That's why a model this size can still respond in seconds: at any given moment, it's only running a small fraction of itself.

This isn't unique to Kimi K3. Mixture-of-experts is the architecture behind most of the largest AI models released in the last two years, from multiple labs. Once you can name the pattern, model announcements that used to sound like arms-race hype start reading as: how many experts, and how many activate at once. That's a durable way to size up any new trillion-parameter headline going forward.

What else K3 does

  • 1 million tokens of context in a single conversation — roughly several complete novels' worth of text, read in one sitting.
  • Native image and video understanding — not bolted on as a separate step.
  • Open weights — no license fee to use the model itself, unlike most frontier-class AI.

The honest caveat

Free describes the license, not the electricity bill. Running the full 2.8-trillion-parameter model yourself takes a rack of serious, expensive hardware — this isn't a laptop download. Most people never need to touch that side of it: Moonshot's own kimi.com gives free-tier chat access to K3 directly, no credit card required.

Why it matters for your AI budget

The pattern to watch isn't this one model — it's the trend line. Frontier-level AI capability keeps showing up in free, open form faster than almost anyone predicted a year ago, and mixture-of-experts is a direct reason it can. If you're planning AI tool spend for next quarter, the bar for good enough, and free just moved up — and it'll keep moving.