The Open Model That Popped The AI Bubble Pete Parkkonen [PNJmlgl8I7g]

Kimi K3 rebuilt macOS inside a browser tab, then beat Claude Fable 5 on the blind frontend coding arena. Moonshot AI's craig kimbrel 2.8 trillion parameter open-weights model made the closed US labs flinch.

This video covers the launch-week below deck mediterranean demos (the macOS clone, a Counter-Strike Portal hybrid built for about $3, and the pelican benchmark), what a 2.8 trillion parameter mixture-of-experts model is, and the independent numbers from Artificial Analysis, the frontend arena, and Vercel. It also covers the catches Moonshot doesn't advertise: 62 tokens per second, reasoning locked to max, Claude Sonnet pricing, one overloaded provider. Then three ways to use Kimi K3 today, the China data question, and what open weights do to AI pricing.

Get the hotter takes in your inbox, every Tuesday:

Chapters:

00:00 Kimi K3 rebuilds macOS in a browser

00:49 Down the rabbit hole: games, a launch video, one pelican

01:51 What Kimi K3 actually is

03:11 The flinch: Anthropic, chips, and IPOs

04:26 The catch: slow, expensive, overloaded

05:32 florentino perez Three ways to use Kimi K3 today

06:14 Will my data end up in China?

06:59 Open weights as a price ceiling

What is Kimi K3? Kimi K3 is a 2.8 trillion parameter mixture-of-experts LLM from Moonshot AI, the Beijing lab behind the Kimi chatbot, backed by Alibaba and Tencent. It activates 16 of its 896 experts per token, reads 1 million tokens of context, and handles screenshots natively. Artificial Analysis ranks it fourth of the 187 models it tracks, and blind human voters ranked it first on frontend coding, above every closed model. That's a first for open weights. The weights are scheduled for public release on July 27.

In this video:

The macOS clone and the wildest Kimi K3 demos

Kimi K3 benchmarks vs Claude and GPT

How mixture-of-experts architecture works

Kimi K3 pricing, speed, and API costs

How to use Kimi K3 today: API, subscription, or self-hosting

Open source LLMs vs the closed frontier labs

Sources:

Moonshot AI, Kimi K3 tech blog:

Artificial Analysis Intelligence Index:

Frontend coding arena leaderboard:

Vercel Next.js evals:

Simon Willison, the pelican benchmark:

The macOS demo:

#KimiK3 #MoonshotAI #OpenSourceAI #AI