*Companion to [Big Tech AI Subscriptions](/publications/big-tech-ai-subscriptions/)
and [If the Big Labs Block You](/publications/if-the-big-labs-block-you/),
which both exclude Chinese developers. This is the piece that doesn't. Prices
as of early October 2026.*

## The short version

Chinese AI is not hard to reach. Independent tracking puts **48 of 57** Chinese
AI products as globally accessible, and the major chat services take a plain
Gmail address — **no Chinese phone number required**. Several are free with no
paid tier at all, and in two categories (open weights, video generation) they
are the leaders rather than the alternative.

## Text and code

| Service | Developer | Sign-up | Free tier | Paid plans | Best at |
|---|---|---|---|---|---|
| [DeepSeek](https://chat.deepseek.com) | DeepSeek | Global email (Gmail/Yahoo) | Yes, no stated message limit | **No consumer tier** — API from $0.30/M (V4.1 Flash), $1.32/M (V4 Pro), half price off-peak | Frontier reasoning at the lowest price anywhere; **MIT-licensed weights** |
| [Qwen Chat](https://chat.qwen.ai) | Alibaba | Global email, no sign-up needed for basic use | Yes, works internationally | Coding plans: Lite **$8** · Standard **$25** · Pro **$80** | The most complete free product — chat, image generation, image/video understanding, documents, voice |
| [Kimi](https://kimi.com/en) | Moonshot AI | Email or phone | Yes | Kimi Code **$19** / **$39** / **$99** / **$199** | Long-context reading and agentic work; the largest open-weight model anywhere (K3, 2.8T) |
| [Z.ai](https://chat.z.ai) | Z.ai (ex-Zhipu) | Email | Yes | GLM Coding Plan ~**$18** Lite · ~**$30–80** Pro/Max (figures vary by source) | Coding and terminal agents; GLM-5.2 is MIT-licensed |
| Doubao | ByteDance | — | — | — | **China-only.** Not accessible internationally |
| Ernie Bot | Baidu | Historically needed a Chinese number | — | — | Treat as China-first; verify before relying on it |

**DeepSeek is the value outlier.** Free chat with no message cap, no consumer
subscription to buy, and API pricing roughly an order of magnitude under
Western frontier models. The weights are MIT — the most permissive licence
attached to anything this capable.

**The coding plans are the real story for developers.** Qwen Lite at $8 and
Z.ai Lite at ~$18 undercut every Western coding subscription, and Kimi Code
tops out at $199 where Western equivalents reach $300–500.

## Images and video

This is where Chinese services lead outright rather than merely compete.

| Service | Developer | Free tier | Paid plans | Notes |
|---|---|---|---|---|
| [Kling](https://klingai.com) | Kuaishou | **66 credits/day, no card** | Standard **$6.99** first month (**$8.80** after) → Ultra **$159.99** | **Kling v3 currently tops the video arena.** Most generous free tier of any major video generator |
| [Hailuo](https://hailuoai.video) | MiniMax | Yes | **$7.99** (1,000 credits) · **$27.99** Pro · **$63.99** Master → **$199.99** | Solid physics, 1080p/24fps |
| [Dreamina](https://dreamina.capcut.com) | ByteDance | Free daily credits | from ~**$15–18** | Consumer route to the Seedance models, rated strongly for commercial work |
| [Qwen-Image](https://chat.qwen.ai) | Alibaba | Via Qwen Chat | Included | Best open image quality — but the **downloadable weights are research-licensed, not commercial** |

For video specifically, the honest ranking puts **Kling v3 first**, ahead of
Western options, with Google's Veo the only non-Chinese model competing at the
top. If you've been avoiding Chinese services on principle, video is where
that costs you the most.

## The four caveats that actually matter

1. **Jurisdiction, not just location.** Your prompts and uploads land with a
   company subject to PRC law. For personal use that may be irrelevant; for
   client data, regulated work or anything under GDPR, it's the whole question.
   This is a compliance decision, not a performance one.
2. **Regional pricing gates.** Kling, Dreamina and Hailuo all gate pricing
   pages by region, so the figures above come from third-party trackers and
   may not match what you're shown. Confirm in-app before subscribing.
3. **Content restrictions are real.** Politically sensitive topics are
   constrained in ways Western services don't constrain them. For code and
   data work you will rarely notice; for research, journalism or policy work
   you will.
4. **Open weights are the escape hatch.** DeepSeek (MIT), GLM-5.2 (MIT) and
   Kimi K2's modified MIT can all be downloaded and run on your own hardware
   — which removes caveats 1 and 3 completely. Note the licences drift: GLM-5.3
   moved to a bespoke licence, Kimi K3 added a revenue-share threshold, and
   Qwen-Image is research-only. Check the exact revision.

## My take

The reflex to exclude Chinese services is usually about data jurisdiction, and
that reflex is sound — for hosted use with real data. What doesn't follow is
excluding the *models*, because the same labs publish the most permissive
weights in the industry. **Run DeepSeek V4 or GLM-5.2 on your own hardware and
the jurisdiction question evaporates**, which is a materially different
proposition from sending your data to Hangzhou.

If you'll use hosted Chinese services: **DeepSeek for reasoning on non-sensitive
work** (free, uncapped, hard to beat), **Qwen Chat as the most complete free
product**, and **Kling for video**, where it's simply the best available. Pay
for the $8–18 coding plans only after checking your own company's data policy,
because that's the constraint that will actually stop you, not capability.

And one thing worth knowing even if you never sign up: **avoiding Chinese
models by choosing a non-Chinese vendor often doesn't work.** Cursor's
Composer 2 is built on Kimi K2.5. Writer's Palmyra X6 is reportedly
post-trained GLM-5.2.

The sharpest example is Japan. **Sakana AI's Namazu models — behind Sakana
Chat, a Japan-first consumer assistant explicitly framed as sovereign AI — are
post-trained on Kimi K2.6 and DeepSeek V3.1.** A national-champion AI product
built on Chinese open weights is the clearest illustration available that
these weights are now infrastructure rather than a competing product. Ask what
the base model was; the answer is often Chinese, and sometimes withheld.

## Further reading

- [Chinese AI models compared: DeepSeek, Qwen, GLM, Kimi](https://geotoolbox.ai/blog/chinese-ai-models-compared)
- [Best Chinese AI tools 2026](https://seeai.net/best/best-chinese-ai-tools/) — accessibility notes per service
- [Open-model coding plans compared](https://www.mindstudio.ai/blog/open-model-coding-plans-glm-kimi-deepseek)
- [Kling pricing](https://www.atlascloud.ai/blog/tips/kling-ai-pricing) · [Hailuo pricing](https://www.atlascloud.ai/blog/tips/hailuo-ai-pricing-cost)
- [Self-Hosting the Frontier](/publications/local-ai-models-2026/) — running these weights yourself