Contents
- 8 Chinese AI models at a glance
- How to choose: 5 axes
- After the answer, sound familiar?
- What you really wanted
- Where Chinese AI stops
- SIMY connects the rest
- One meeting, compared
- By industry: Chinese AI and what SIMY delivers
- Get started in 3 steps
- Chinese AI explained: local use and safety
- FAQ
- Sources
8 Chinese AI models (LLMs) at a glance
Eight companies, the same seven points (as of October 2026). Open weights means the model's weights are published, so you can run it yourself. An API is the paid gateway for calling the model from your own tools.
DeepSeek
DeepSeek (Hangzhou, China)
- Chat: Web and app. Free
- API from outside China: Yes. OpenAI- and Anthropic-compatible
- Open weights: V4.1-Flash and others, MIT license
- On your computer: Older models such as small R1 versions. Latest is cloud-only in Ollama
- Where data is processed: Says it's stored on servers in China
- Languages: No official multilingual claim. Works in Japanese and other languages in practice
- Official action: Warnings from Japan's Personal Information Protection Commission and the government (Digital Society Promotion Council executive committee). Restricted in other countries too
Qwen (通义千问)
Alibaba Group (Hangzhou, China)
- Chat: Qwen Chat and app. Free
- API from outside China: Model Studio. Choose regions such as Tokyo
- Open weights: Qwen3.8-27B under Apache 2.0. Largest models under a custom license
- On your computer: Yes. qwen3.8 (27B) is about 18 GB
- Where data is processed: Qwen Chat says Singapore. API lets you choose Tokyo, among others
- Languages: Claims 119 languages and dialects (Qwen3 generation)
- Official action: Included in Taiwan's inspection of Chinese AI apps (reported)
Kimi
Moonshot AI (Beijing, China)
- Chat: kimi.com and app. Free tier and paid membership
- API from outside China: Yes (Singapore entity). OpenAI- and Anthropic-compatible
- Open weights: Kimi K3 under a custom license (conditions such as revenue size)
- On your computer: No. Cloud-only in Ollama
- Where data is processed: China for the China-facing app, Singapore for the API
- Languages: No official multilingual claim
- Official action: No named measures found
GLM (Z.ai)
Zhipu AI (Beijing, China)
- Chat: chat.z.ai. Free
- API from outside China: Z.ai (Singapore entity). Anthropic-compatible, GLM Coding Plan
- Open weights: GLM-5.3 under a custom license, GLM-5.3-Flash under MIT
- On your computer: glm-4.7-flash (about 19 GB) works. Latest is cloud
- Where data is processed: International service says Singapore, in principle
- Languages: No official multilingual claim (English and Chinese)
- Official action: Added to the US Commerce Department's export control list (January 2025, reported)
MiniMax
MiniMax (Shanghai, China)
- Chat: Apps such as MiniMax Code (formerly MiniMax Agent)
- API from outside China: Yes. Anthropic-compatible recommended, OpenAI-compatible too
- Open weights: MiniMax-M3 under a custom license (commercial use requires attribution, notice and more)
- On your computer: Not for personal computers. Cloud in Ollama
- Where data is processed: International app says Singapore, in principle
- Languages: No official multilingual claim
- Official action: No named measures found
Doubao (豆包)
ByteDance (Beijing, China)
- Chat: Doubao (豆包), for China
- API from outside China: BytePlus ModelArk
- Open weights: Flagship is closed. Seed-OSS-36B under Apache 2.0
- On your computer: Not in Ollama's official library
- Where data is processed: Not confirmed officially
- Languages: No official multilingual claim
- Official action: Included in Taiwan's inspection of Chinese AI apps (reported)
Hy (formerly Hunyuan)
Tencent (Shenzhen, China)
- Chat: Yuanbao (元宝), for China
- API from outside China: Available via Tencent Cloud TokenHub
- Open weights: Hy4 preview and Hy3 under Apache 2.0
- On your computer: Not in Ollama's official library. Built for GPU servers
- Where data is processed: Not confirmed officially
- Languages: No official multilingual claim
- Official action: Included in Taiwan's inspection of Chinese AI apps (Yuanbao, reported)
ERNIE (文心一言)
Baidu (Beijing, China)
- Chat: ERNIE Bot (文心一言), for China
- API from outside China: Qianfan. Mainly for the Chinese market
- Open weights: ERNIE 4.5 series under Apache 2.0. 5.x series is closed
- On your computer: Not in Ollama's official library (community builds exist)
- Where data is processed: Not confirmed officially
- Languages: No official multilingual claim
- Official action: Included in Taiwan's inspection of Chinese AI apps (reported)
How to read this: "Not confirmed officially" means we couldn't verify it from official sources, not that it doesn't exist. Prices change often, so we don't list amounts; links to the official pricing pages are in the table below.
How to choose Chinese AI: price, local use, languages and safety
"Which is smartest" changes every month. Choose on five axes that don't, and you won't get lost.
- 1
By API price
DeepSeek (V4.1-Flash) and MiniMax (M3) are especially cheap per token. DeepSeek charges double on weekdays from 9:00–12:00 and 14:00–18:00 Beijing time (01:00–04:00 and 06:00–10:00 UTC), except Chinese public holidays. If you code with it every month, compare flat-rate options such as the GLM Coding Plan too.
- 2
By open weights
The least restrictive are MIT (DeepSeek V4.1-Flash, GLM-5.3-Flash) and Apache 2.0 (Qwen3.8-27B, Hy4 preview and Hy3, ERNIE 4.5). Kimi K3, GLM-5.3 and MiniMax-M3 come with conditions, so read the license before commercial use.
- 3
By local use (your own computer)
Realistic options are Qwen3.8-27B (about 18 GB), GLM-4.7-Flash (about 19 GB) and small DeepSeek-R1 versions. Each company's latest flagship is huge, and even in Ollama it carries the cloud tag (it runs on external servers).
- 4
By language
Only Qwen officially lists broad language support (119 languages and dialects). The others often work in other languages too, but with no official guarantee, so test them on your real work documents.
- 5
Safety: by where your data goes
Even within one company, where your data goes, and which country's laws apply, depends on the entry point. If no data may leave your hands, only running open weights in your own environment fits.
After comparing Chinese AI, sound familiar?
The comparison and the prototype took seconds with a cheap AI. Here's what happens next.
The comparison looked great. The owner and deadline got lost in the meeting.
New models every month. What you adopted and why is buried in chats.
Someone was going to check data handling with IT. Did anyone ask?
Any AI gives you an answer in seconds. Whether work moves is decided after the answer.
What you really wanted after the AI answer
Same meeting. Same AI. The moment the answer arrived, you wanted this.
Research and drafts stay with your usual AI. Decisions, owners and deadlines are already lined up.
SIMY flags work that's waiting or nearly due before your next meeting.
What you adopted and what's next stay together, across models and chats.
"Which model do we trial?" That's the one question you get back.
This starts from the next meeting after you connect your meetings and chats to SIMY. Keep using Chinese AI for what it does best.
This is where Chinese AI answers stop
Every model is good at answering what you ask. After that, nobody picks it up.
- QuestionYou ask
- Research, translationCheap and fast
- Drafts, codeIn bulk
- Owners, deadlines
- Report
- Internal update
- Follow-up checks
* Flow shown for illustration only
Chat, API or local, it stops in the same place. Switch models all you like: you still do the asking every time, and you still check whether it got done.
SIMY connects the rest. Keep your AI, hand off what comes after
SIMY fills the four blanks. Only one box comes back to you: the decision.
- Question
- Research, translation
- Drafts, code
- Owners, deadlinesPicked up from meetings
- ReportDrafted in Gmail
- Internal updatePosts once you approve
- DecisionOne question back
* Flow shown for illustration only
- Picks up from meetings and chat
- Does it your way
- Returns only decisions
SIMY finds the follow-up work in your meeting records and chats on its own. It learns the steps and checks you repeat from your conversations, follows the same approach, and lists what's done in My Actions.
- Sign up for SIMY and connect ChatGPT (Codex)SIMY's cloud runs use your own ChatGPT account. Codex usage falls within your ChatGPT plan.
- Install the SIMY desktop appSIMY can use Claude Code on your computer to do the work. It turns repeat work in your Claude Code and Codex chats into "Suggestions from SIMY." Chat text is not stored on our servers. Download here
- Connect your meeting recordsConnect Plaud or Notion on the integrations screen. Your meeting records become the starting point SIMY picks work up from.
* SIMY screen shown for illustration only
How SIMY relates to Chinese AI: At this time, SIMY does not connect to the LLMs on this page. SIMY itself runs on your ChatGPT (Codex) account and, through the SIMY desktop app, on Claude Code on your computer. Leave research, translation and drafts to each model, and SIMY takes on the follow-through for what the meeting decided.
One meeting, compared: the AI's notes and SIMY's next steps
A sample from a 50-minute review to pick which AI to trial in-house, organized with AI. The AI's comparison notes on the left, SIMY on the right.
- Options
- DeepSeek for cheap API, Qwen for language support and local use, GLM for coding.
- Issues
- How much internal data can be sent out. Cost of running it in-house.
- To-dos
- Check with IT. Trial on the same task and estimate costs.
- Due
- By the next review (October 20).
Check with IT: you, Oct 8. Trial on 3 models: Engineering, Oct 15. Cost estimate: Planning, Oct 17.
Models to trial, the kinds of data you plan to send and where each company processes data, in your usual tone. Saved as a Gmail draft.
"We'll trial with public information only. Results at the review on the 20th." Posted to chat once you approve.
Trial via API, or run it in-house? Only what needs your decision comes back to you.
AI compares. SIMY moves. You still make the call.
By industry: how teams use Chinese AI, and what SIMY delivers
What SIMY prepares once Chinese AI has given you the answer. Solo or as a team, the pattern is the same.
Engineering Kimi, GLM: coding via compatible APIs
- Design review meeting→Agreed fixes and owners
- After writing code→Next task in Claude Code on your PC
- End of the week→Progress and pending decisions
Trade and sourcing Qwen: Chinese translation and summaries
- Supplier check-in→Owners, deadlines, reply draft
- Price or lead-time change→Internal update, one decision
- Before a reply deadline→Reminder on work awaiting replies
Marketing and content DeepSeek, MiniMax: volume on cheap APIs
- Planning meeting→Chosen ideas, owners, deadlines
- Draft review→Request listing the edits
- Before publishing→Pending checks, one decision
IT and operations Setting AI usage rules
- AI policy meeting→Agreed rules, who announces them
- Questions from teams→Reply drafts, pending checks
- Quarterly review→Key issues, one decision
Get started in 3 steps, AI unchanged
- Use Chinese AI as usualResearch, translation, drafts, code. Chat, API or local. Follow your company's rules on what you enter.
- Sign up for SIMY and connect ChatGPTCloud runs use your ChatGPT (Codex) account. If you use Claude Code, install the SIMY desktop app too. See how it connects above.
- Finish one meetingOwners and deadlines, a report draft and the update text arrive. Decide and send. That's it.
How sign-up works: choose a plan → verify your email → pay by card. There is no free plan. See pricing. Get the desktop app from the download page.
How you use AI doesn't change. SIMY doesn't replace Chinese AI. It's a separate system that takes on the work after the answer.
Chinese AI explained: comparison table, local use, compatible APIs, safety and regulation (in depth)
Open these when you want to check how things work or what to watch out for. Model names and specs are as of October 2026.
What is Chinese AI, and why is it cheap and smart?Open weights and price competition set it apart
"Chinese AI" refers to large language models (LLMs) built by Chinese companies and research institutes, and the chat services built on them. The release of DeepSeek-R1 in January 2025 (the so-called DeepSeek shock) brought them to wide attention, including in Japan.
Three traits
- Lots of open weights: Many companies publish their model weights so anyone can run them in their own environment. Licenses are a mix, though: MIT, Apache 2.0 and custom licenses with conditions.
- Cheap APIs: Techniques such as running only part of the model for each query (MoE, Mixture of Experts) cut computation and lower API prices. DeepSeek-V4.1-Flash, for example, has 552B parameters in total, but only 8B to 16B are active at inference time.
- Fast turnover: July to September 2026 alone brought Kimi K3, MiniMax-M3, GLM-5.3, DeepSeek-V4.1-Flash and Hy4 preview. Remembering models by their role, not their name, makes them easier to keep up with.
Many companies also offer compatible APIs so you can use their models from developer tools such as Claude Code and Codex (see "Compatible APIs" below).
Comparison table: 8 Chinese AI modelsThe cards above, side by side in one table
Scroll the table sideways →
| AI | Main models | Latest open weights | Run locally with Ollama | Anthropic-compatible API | Official multilingual claim |
|---|---|---|---|---|---|
| DeepSeek | V4.1-Flash, V4-Pro | Yes (MIT) | Older models only (R1 and others) | Yes | No |
| Qwen | Qwen3.8-Max, Qwen3.8-27B | Yes (27B under Apache 2.0) | qwen3.8 (27B) | Via Coding Plan | Yes (119 languages and dialects) |
| Kimi | Kimi K3 | Yes (custom license) | No (cloud only) | Yes | No |
| GLM (Z.ai) | GLM-5.3, GLM-5.3-Flash | Yes (5.3 custom, Flash MIT) | glm-4.7-flash | Yes | No |
| MiniMax | MiniMax-M3 | Yes (custom license) | No (cloud only) | Yes (recommended) | No |
| Doubao | Doubao-Seed series | Older models only (Apache 2.0) | No | Not confirmed officially | No |
| Hy (formerly Hunyuan) | Hy4 preview, Hy3 | Yes (Apache 2.0) | No | Not confirmed officially | No |
| ERNIE | ERNIE 5 series | 4.5 series only (Apache 2.0) | No (community builds exist) | Not confirmed officially | No |
Check prices on each company's official page: DeepSeek, Qwen, Kimi, Z.ai, MiniMax.
In Ollama doesn't mean it runs locally (the cloud tag)"Available in Ollama" is not the same as "local"
Ollama is the go-to tool for running LLMs on your own computer. But its official library also lists models with a "cloud" tag, which run on Ollama's cloud instead of your machine. As of October 2026, the latest flagships from DeepSeek, Kimi, GLM and MiniMax are cloud-tag only.
Scroll the table sideways →
| Type | Ollama models (examples) | Where it runs |
|---|---|---|
| Runs locally | qwen3.8 (27B, about 18 GB), glm-4.7-flash (about 19 GB), small deepseek-r1 versions (7B is about 4.7 GB) | On your computer |
| Cloud only | deepseek-v4.1-flash, deepseek-v4-pro, kimi-k3, glm-5.3, minimax-m3 | Ollama's cloud (external servers) |
| Not in the library | Doubao-Seed series, Hy, latest ERNIE | Your own server, from Hugging Face with vLLM or similar |
How to tell
- Check the model page: If the tag or size column in Ollama's library says "cloud," it won't run on your computer.
- Check the download size: Local models pull a file of several GB to tens of GB onto your computer. If nothing downloads and it works right away, it's cloud.
- Check your goal: If the point is to keep data in-house, cloud-tag models don't qualify.
For setup steps and memory guidelines, see How to run Qwen locally (Ollama, LM Studio).
Anthropic- and OpenAI-compatible APIs: use them from Claude Code or CodexJust switch the endpoint, but your data goes somewhere new too
Many Chinese AI providers offer "compatible APIs" you can call the same way as the OpenAI or Anthropic APIs. Point a developer tool such as Claude Code or Codex at a different endpoint, and you swap only the model.
Scroll the table sideways →
| AI | Anthropic-compatible (Claude Code and others) | OpenAI-compatible (Codex and others) |
|---|---|---|
| DeepSeek | https://api.deepseek.com/anthropic | https://api.deepseek.com. Officially states support for the Responses API and Codex |
| Kimi | https://api.moonshot.ai/anthropic | Yes |
| GLM (Z.ai) | https://api.z.ai/api/anthropic. Flat rate with GLM Coding Plan | Yes |
| MiniMax | Anthropic SDK compatibility recommended | Chat Completions and Responses API compatible |
| Qwen | Works with various tools via Coding Plan | Model Studio is OpenAI-compatible |
Check before you use it
- Your code is sent too: Once you switch the endpoint, not just your prompts but the code and file contents your tool reads are sent to that company's servers.
- Policies and contracts: Whether you may send internal source code to an external API depends on your employer's policies and your contracts with clients.
- Key management: Pass API keys through environment variables, and never commit them to a repository.
How this relates to SIMY: Switching endpoints is a setting in Claude Code or Codex, and SIMY doesn't provide guidance or support for it. At this time, SIMY does not connect to the LLMs on this page.
Safety and data handling: Japanese regulators' warnings and moves abroadNot a ban, but a call to judge the risks
Each entry point sends data somewhere different
- DeepSeek: Says data collected through both the app and the API is stored on servers in China. You can opt out of training use in settings.
- Kimi: Says data is stored on servers in China for the China-facing app, and in Singapore for the API.
- Z.ai (GLM), MiniMax: International services are run by Singapore entities and say data is processed in Singapore, in principle (MiniMax's API also mentions a US data center).
- Qwen: Qwen Chat says data is processed in Singapore. The API lets you choose regions such as Tokyo.
- Running open weights yourself: Data never leaves your own environment.
Warnings from Japanese authorities
- Japan's Personal Information Protection Commission (February 3, 2025, updated March 5): Warned users that DeepSeek stores data on servers in China and that Chinese law applies to it.
- Executive committee of Japan's Digital Society Promotion Council (February 6, 2025, reported): Under the name of this council, whose secretariat is Japan's Digital Agency, an advisory on using DeepSeek and similar generative AI for government work was issued to government agencies. It doesn't ban them outright; it asks agencies to decide based on the risks.
As of October 2026, we found no law or order in Japan that bans individuals from using Chinese AI outright. If your employer or clients have their own usage rules, follow those.
Moves abroad (reported)
- DeepSeek: Italy's authority blocked the app, Australia banned it on government devices, South Korea suspended new downloads, and a German authority asked app stores to remove it, among others (2025).
- Zhipu (GLM): Added to the US Commerce Department's export control list (the Entity List) in January 2025.
- Chinese AI apps in general: In November 2025, Taiwan's National Security Bureau published inspection results for DeepSeek, Doubao, ERNIE Bot, Tongyi and Yuanbao.
Moves abroad are summarized from news reports. Check each authority's announcements for the latest. SIMY's own data handling is published on our Security page and in our Privacy Policy.
Individual guides: DeepSeek, Qwen, Kimi, GLM, MiniMax, DoubaoHow to use them, APIs, local use and data handling
- What is DeepSeek? How to use it, API, languages and safety
- What is Qwen? How to use it, API, local use and safety
- How to run Qwen locally
- What is Kimi? How to use it and Kimi Code
- What is GLM (Z.ai)? How to use it and the GLM Coding Plan
- What is MiniMax? How to use it and its API
- What is Doubao (豆包)? ByteDance's AI
Frequently asked questions
What Chinese AI models (LLMs) are there?
The main ones are DeepSeek, Alibaba Group's Qwen, Moonshot AI's Kimi, Zhipu's GLM (Z.ai), MiniMax, ByteDance's Doubao, Tencent's Hy (formerly Hunyuan) and Baidu's ERNIE. The comparison cards on this page sum up how they differ as of October 2026.
Can I use Chinese AI for free?
Most of the chat apps are free. APIs are pay-as-you-go, and open-weight models cost nothing to use if you run them in your own environment.
Which Chinese AI has the cheapest API?
As of October 2026, DeepSeek (V4.1-Flash) and MiniMax (M3) are especially cheap per token. Rates vary by time of day and input length, so check each company's official pricing page.
Do Chinese AI models support languages other than English and Chinese?
Qwen is the one that officially lists broad language support (119 languages and dialects). Other models often work in other languages too, but there's no official guarantee, so test them on your own work documents.
Can I run Chinese AI models on my own computer?
Some of them. With Ollama, Qwen3.8-27B (about 18 GB), GLM-4.7-Flash (about 19 GB) and small DeepSeek-R1 versions are realistic. Each company's latest flagship is too large for a personal computer.
Does every model in Ollama run on my own computer?
No. Models with the cloud tag run on Ollama's cloud, not on your computer. As of October 2026, DeepSeek-V4.1-Flash, Kimi K3, GLM-5.3 and MiniMax-M3 are cloud-tag only.
Can I use Chinese AI from Claude Code or Codex?
Yes. DeepSeek, Kimi, GLM (Z.ai), MiniMax and others offer Anthropic- or OpenAI-compatible APIs, so you can switch the endpoint and use them. Note that the code your tool reads is then sent to that company's servers too.
Where is the data I enter into Chinese AI stored?
It depends on the company and the entry point. DeepSeek says China; Kimi says China for its China-facing app and Singapore for its API; the international services of Z.ai and MiniMax, and Qwen Chat, say Singapore. If you don't want data to leave your hands, run an open-weight model in your own environment.
Is using Chinese AI banned in Japan?
As of October 2026, we found no law or order that bans it outright. However, in February 2025 Japan's Personal Information Protection Commission and the executive committee of the government's Digital Society Promotion Council issued warnings about DeepSeek and similar services, asking for decisions based on the risks.
Is Chinese AI safe?
It isn't simply safe or unsafe; judge it by which entry point you use and which country stores your data. Keep confidential information out of consumer apps, check the setting that stops your data being used for training, and run open weights in your own environment if data must not leave it.
Can I use open-weight models freely for commercial work?
It depends on the model. MIT and Apache 2.0 models are relatively free to use, but Kimi K3, GLM-5.3 and MiniMax-M3 have custom licenses with conditions such as revenue size or attribution requirements. Read the license text before commercial use.
Does SIMY run on Chinese AI?
No. At this time, SIMY does not connect to the Chinese AI models on this page. It runs on your ChatGPT (Codex) account and, through the SIMY desktop app, on Claude Code on your computer. You can keep using Chinese AI for research and drafts, and leave the follow-through on decisions to SIMY.
Do I have to ask SIMY every time to handle the work after an AI answer?
No. SIMY picks up decisions, owners and deadlines from your meeting records and chats on its own. It learns your way of working from your conversations and returns only what needs your decision.
Sources
Each company's models, licenses, API compatibility and data handling, and the warnings from Japanese authorities, were checked against the official sources below (checked October 1, 2026). Moves abroad draw on news reports. Model names and specs change often, so check each official page for the latest.
- DeepSeek API docs, DeepSeek Privacy Policy, DeepSeek-V4.1-Flash (Hugging Face)
- Qwen3.8-Max (Alibaba Cloud Model Studio), Qwen3.8-27B (Hugging Face)
- Kimi K3 (Hugging Face), Kimi Open Platform Privacy Policy, Kimi data usage
- GLM-5.3 (Hugging Face), Z.ai: using it with Claude Code, Z.ai Privacy Policy
- MiniMax-M3 (Hugging Face), MiniMax local deployment guide, MiniMax Privacy Policy
- ByteDance Seed (Hugging Face)
- Hy4 preview (Hugging Face), Hy3 available globally (Tencent)
- Baidu ERNIE (Hugging Face)
- Ollama library (DeepSeek), glm-4.7-flash, qwen3.8
- Information on DeepSeek (Japan's Personal Information Protection Commission, February 3, 2025; in Japanese), Report on the advisory to government agencies (INTERNET Watch; in Japanese)
- How countries have responded to DeepSeek (Al Jazeera), Zhipu added to the Entity List (Bloomberg), Taiwan National Security Bureau inspection (Focus Taiwan)