The Chinese models (GLM, Qwen, Seedance, Kling)

Z.ai renamed itself from Zhipu and listed on the Hong Kong stock exchange - the first AI lab in the world to do so. It makes GLM, a family of models that ships a new version roughly every two to three months, while the Western labs go a whole year between releases. The latest version is open source under the MIT license, the most permissive kind - you take the code and do whatever you want with it, even sell it. No locked door, no secret recipe behind it.

Alibaba stands behind Qwen and overtook Llama in downloads - the open model from Meta, the parent company of Facebook. It's not charity. Anyone who downloads Qwen and runs it on a server sooner or later pays for the server itself, and the cloud is theirs. The model is the bait. The business is the infrastructure underneath.

ByteDance makes Seedance, a video model. You write text, upload a photo, it returns a clip with synced sound, music and speech in one motion. It ranked first in almost every international video-model leaderboard right from the start, ahead of the American competitors.

Kling is the work of Kuaishou, a Chinese short-video platform, a domestic competitor to TikTok. The latest version produces realistic motion, synced sound in several languages and dialects. It's no longer a toy for fun - ad agencies use it for real production.

The pattern isn't a coincidence. These companies find it hard to sell directly to the West because of sanctions and distrust, so they give the models away and profit in a roundabout way - through cloud, chips, infrastructure. The result is that over sixty percent of requests to models through OpenRouter (a marketplace where apps buy access to different AI models) now go through Chinese systems. Meta dropped completely out of the open-model leaderboard - the one it topped only a short while ago.

An example from this week is DeepSeek. On October 17, 2026 the lab released V4.1-Flash-1017 with the weights under MIT, and in the same announcement opened its Search & Fetch API to everyone: 0.75 dollars per 1,000 requests per endpoint. Through the server tools in DeepSeek Harness, Claude Code and Codex it is free for DeepSeek API users until payment starts on October 31, 2026. The model is given away, the search is sold.

They opened the door we closed on them - and more traffic now flows through it than through our own.

Here's what's next

While we argue over whether to pay for yet another subscription to a closed model, somewhere in Hangzhou and Beijing they're putting the code online for free and waiting for the world to download it itself. Not because they're generous - because the closed door left them no other choice. It turned out to be a good strategy.

A parallel world is growing that doesn't ask our permission to exist and that we don't notice until the numbers get uncomfortable. I'm not afraid of it. I just stop pretending it isn't there. If the best free model tomorrow is Chinese, and I'm paying for an American one purely out of habit - that's my problem, not theirs.

The visual is generated code-art, no third-party images.