In Class 10 you learned to route across providers through OpenRouter. You could pick OpenAI, Anthropic, or Google and fall back when one was slow or expensive. That is a good start, but it leaves half the market off your short list.

Some of the strongest coding models, the cheapest API prices, and the most useful open-weight releases right now come from Chinese labs. If you only compare three American brands, you will overpay for some tasks and miss the best tool for others.

By the end of this lesson, you can name the five major Chinese model families, explain why builders care about them, and separate model origin from hosting, API provider, and data processing.

Meet the five families

You do not need to memorize every release. You need to recognize the family name when you see it on OpenRouter, Hugging Face, or a leaderboard, and know what to check next.

Release names change every few months. Qwen3.x, DeepSeek V4 Flash/Pro, GLM-5.x, MiniMax M3, and Kimi K2.7/K3 were the visible generations in mid-2026 (note: "GPL" in older notes usually means GLM from Zhipu AI). Do not anchor on the version number. Anchor on the habit: when a new version appears, evaluate it the same way you evaluate any other model.

And do not stop at China vs the US. Meta belongs in the same comparison. Its Llama family (Llama 3.1/3.3, Llama 4 Scout and Maverick) are the most widely deployed open-weight models in the world — free weights you can self-host or call cheaply through providers (often $0.05–$0.20 input per million for smaller sizes). For everyday copywriting, blog drafts, product descriptions, and rewrites, Llama 3.1 70B and Maverick are outstanding value: fluent, steerable, and cheap enough to iterate. Meta's newer Muse family pushes higher: Muse Spark 1.3 (Sept 2026, 1M context, ~$1.25 input / $4.25 output, or ~$0.10 / $0.20 on its contributor tier) scores near the top for coding and long agentic work and also ranks very highly for writing and instruction-following — a strong example of a non-ChatGPT, non-Claude default worth testing for copy and code alike.

Why builders care: four practical reasons

1. Capability. Qwen and DeepSeek coding models regularly sit near the top of public coding and reasoning comparisons alongside American frontier models. GLM and Kimi models are competitive on long-document and agentic tasks. For code generation, refactoring, and tool use, these are not curiosity picks. They are short-list picks.

2. Economics — much cheaper, and more than good enough. This is the point that changes budgets. On public pricing in 2026, a flagship US model can list around $3–$5 per million input tokens and $15–$25 per million output tokens, while DeepSeek V4 Flash lists around $0.14–$0.22 input and $0.28–$0.66 output, Qwen's efficient tiers list as low as $0.03–$0.15 input, and GLM routes are in the same low band. That is not a 20% discount — it is often 10–25x cheaper per token for routine work. For summarization, classification, first drafts, bulk rewrites, and many coding assists, these models are not just "cheap for what you get." They are fully adequate for the job. Check current prices on OpenRouter before you lock in any number, but the pattern is durable: do not pay flagship prices for non-flagship difficulty.

That is why the rule from Class 08 matters so much here: use the right model for the right job instead of getting stuck on ChatGPT or Claude for everything. A strong default for hard reasoning, a cheap Chinese-family model for high-volume routine work, and a specialized model where it fits will beat single-brand loyalty on both quality and cost.

3. Context windows. Kimi made its name on very long context. Qwen and GLM releases commonly ship with large context windows as well. When your task is "read this whole repository" or "read these fifty pages," context size is part of task fit, which you learned in Class 08.

4. Open weights. Qwen, DeepSeek, GLM, and MiniMax all publish open-weight models on Hugging Face. That means you can inspect the license, download the weights, run them locally or on your own server, or access the same weights through a hosted provider. Open weights do not mean free or private by default — you still pay for hardware or API calls — but they give you hosting choices a closed model never gives you.

More choice also means faster competition. When five more labs ship good models, every provider has to improve quality, cut prices, or open larger contexts to keep your business.

The distinction that prevents confusion

This is the most important section in this lesson. Four things that sound like one thing are actually four separate decisions:

Model origin       → who trained the weights (example: Alibaba trained Qwen)
Weights hosting    → where the downloadable files live (example: Hugging Face)
API provider       → who runs the servers that answer your request (example: OpenRouter route, Alibaba Cloud, a US cloud host)
Data processing    → where your prompt and files are actually processed and under what policy

A Qwen model can be served by a non-Chinese cloud provider. A DeepSeek model can be served through OpenRouter to a route you choose explicitly. Your prompt is processed wherever the API provider you selected runs it, under that provider's data policy — not automatically in the country where the model was trained.

So when someone says "that is a Chinese model," ask three follow-up questions before you make any decision:

1. Which weights, exactly, and under what license? 2. Which provider is serving my request, and where does that provider process data? 3. What does that provider's data-retention policy say?

Class 10 taught you the provider layer. This lesson applies it: origin tells you who built the capability. The provider and its policy tell you where your data goes.

A practical way to think about them

Do not sort models by nationality. Sort them the way Class 08 taught you: by task fit.

Exercise: compare two families on facts, not headlines

Pick two families — for example, Qwen and DeepSeek, or GLM and Kimi.

For each one, write one short paragraph covering:

• One known strength relevant to your work (coding, reasoning, long context, speech)
• Current API pricing for one comparable size, from OpenRouter or the official provider page
• Whether open weights exist and under what license
• Two access paths (for example: official API and an OpenRouter route, or Hugging Face weights plus a local run)

Finish line: you have a half-page note with four facts per family. No adjectives about countries. Just task fit, economics, license, and hosting.

Verify: can you answer "if I call this model through this provider, where is my prompt processed?" If not, you have not finished the provider step.

Common failure mode: comparing version numbers instead of comparable releases — a small Qwen against a flagship Claude, for example. Compare models aimed at the same job.

Check your understanding

1. Name the five Chinese model families and the organization behind each. 2. What are the four practical reasons a builder should evaluate them? 3. What is the difference between model origin, weights hosting, API provider, and data processing? 4. Why is "it is a Chinese model" not enough information to decide where your data goes? 5. For your current project, which task would you test first on a Qwen, DeepSeek, or GLM model — and why?

Next, you will learn to investigate any of these releases before you trust it. Lesson 11.2 turns Hugging Face into your research desk: org page, model card, license, size, quants, and hardware.