Thesis
The most useful way to track China AI models in English is not a giant benchmark sheet. It is a compact weekly list of the labs and model families that keep changing builder decisions, paired with the English-accessible places where you can verify what actually shipped.
Decision in 20 seconds
Use this page if your main question is which China AI models and labs belong in your watchlist. Start with DeepSeek, Qwen, Kimi, MiniMax, ERNIE, Doubao, GLM, and Hunyuan. Use this page to keep the list practical, then verify releases through GitHub, Hugging Face, technical reports, and official docs before anything becomes a roadmap decision.
Who this is for
- Builders and PMs who need a compact weekly watchlist rather than a giant market map.
- English-first teams who want to know which China AI names still matter after the news cycle fades.
- Researchers and evaluators who need a short list of families worth checking for benchmark, access, or license changes.
Who this is not for
- Readers looking for a benchmark leaderboard with one universal ranking.
- People who want every China AI lab listed regardless of builder relevance.
- Readers whose main question is source selection or workflow rather than model families to keep in view.
Use this page when
This page answers the tracker question: which China AI models and labs should I keep on my weekly watchlist in English? If your question is how to run the weekly review, use the workflow guide. If your question is which sites and source types to use, go to the Best Sites page. If your question is what translation lag to expect, use the supporting article.
What is the best way to use a China AI models list?
The best way to use a China AI models list is as a weekly tracking layer, not as a final ranking. Keep a short set of labs and model families in view, watch the English-accessible release channels for each one, and verify benchmark source, API access, and license before you act. This works better than following generic AI news because it keeps the China AI watchlist small, current, and connected to practical verification steps.
How to read this page
- Current watchlist tells you which families deserve a permanent slot in a weekly review.
- Trigger action tells you what kind of change should move an item from "notice" to "review this week."
- Verification links tell you where to confirm the release in English before you repeat or act on the claim.
- Tier 2 names are worth adding only when your scope expands beyond the core builder watchlist.
Current watchlist
| Model or family | Lab or company | Why it stays on the list | What should trigger action | Best English-accessible verification links | RadarAI note |
|---|---|---|---|---|---|
| DeepSeek-V3 / DeepSeek-R1 | DeepSeek | Often resets open-model cost-performance conversations and benchmark comparisons. | New flagship release, benchmark jump, API pricing change, or license shift. | GitHub, Hugging Face, technical report, official docs | Usually one of the first China-origin model families that changes builder evaluation queues. |
| Qwen family | Alibaba Cloud | Frequent releases across sizes, modalities, and OSS-friendly distribution channels. | New family branch, stronger reasoning variant, OSS release, or access update. | QwenLM GitHub, Hugging Face, official docs, technical report | Useful when you want both strong open models and clear English-facing release materials. |
| Kimi family | Moonshot AI | Product-facing launches often shape how people talk about China AI reasoning and UX. | Major product release, reasoning improvement claim, or broader English-facing rollout. | Official product pages, release notes, research posts, English coverage | Worth tracking when the signal is product experience or launch momentum rather than a repo-first release. |
| MiniMax family | MiniMax | Strong candidate when multimodal packaging and practical product access matter more than one benchmark line. | Multimodal launch, API availability change, or pricing / packaging update. | Official docs, release pages, research posts, English summaries | Useful when the question is not just model quality but also product packaging and practical access. |
| ERNIE family | Baidu | Enterprise packaging, cloud distribution, and China-market context can matter more than raw model buzz. | Enterprise release, API or cloud packaging change, or region-access signal. | Official docs, Baidu AI Cloud updates, product pages, English reporting | Important when your decisions depend on enterprise packaging, cloud access, or China-market context. |
| Doubao family | ByteDance | Fast product iteration and ecosystem moves can matter even when the repo story is weaker. | Major product feature release, model refresh, or platform integration move. | Official product pages, research posts, GitHub when available, English summaries | Track when you care about fast product iteration and ecosystem-level movement, not just one repo. |
| GLM family | Zhipu AI | Relevant when you want another strong commercial and API-facing line beyond DeepSeek and Qwen. | New GLM generation, API availability expansion, or enterprise partnership signal. | Official docs, release notes, model pages, English reporting | Add early if your team compares commercial APIs, not just open weights. |
| Hunyuan family | Tencent | Matters when ecosystem distribution, cloud reach, and platform packaging are part of the evaluation. | Cloud release, enterprise access update, or notable multimodal / agent capability move. | Official docs, Tencent Cloud updates, product pages, English summaries | Most relevant when you care about platform leverage and enterprise distribution, not only benchmark chatter. |
Tier 2 names to add only if your scope expands
| Name or family | Why add later | Add it when |
|---|---|---|
| THUDM research line | Strong academic and open research signal, but not always the first builder decision layer. | You care about frontier research repos, not only deployable product choices. |
| SenseNova | Useful for enterprise and multimodal tracking, but less central for a compact weekly builder watchlist. | You need broader enterprise AI vendor coverage inside China. |
| Step family | Worth watching for momentum and ecosystem chatter, but not always a top-8 must-track family. | You start seeing repeated product relevance or customer questions around it. |
| Yi family | Still useful historically and for selective OSS comparison, but less often the first family that changes this week's decision. | Your stack or benchmarks still compare against earlier open-model baselines. |
What should trigger action this week
| If this changes | Why it matters | What to do next |
|---|---|---|
| New flagship model or major version | Could change benchmark comparisons, evaluation backlog, or product positioning. | Read the technical report or model card, then compare against your current default model. |
| API access opens or changes | A model moves from "interesting" to "testable" only when your team can actually use it. | Check docs, pricing, account requirements, and region access before adding it to testing. |
| License terms change | Commercial use assumptions break fast when the license changes across versions. | Read the LICENSE file and model card before sharing a recommendation internally. |
| Benchmark claim gets third-party confirmation | That is often the moment hype becomes evaluation-worthy. | Move the model from watchlist into a short benchmark or prompt test. |
| Distribution or cloud packaging changes | Enterprise and production relevance often depends more on packaging than on one raw score. | Re-check whether procurement, deployment, or regional access just became easier. |
What to verify for every tracked model
| Field | Why it matters | Good source |
|---|---|---|
| Benchmark source | Separates self-reported claims from reproducible evidence | Technical reports, model cards, third-party leaderboards |
| API access | Determines whether your team can actually test the model | Official docs, pricing page, onboarding or account requirements |
| License terms | Determines whether commercial use is allowed or restricted | LICENSE file, model card, official release page |
| Release channel | Shows whether the claim comes from a primary source or commentary only | GitHub repo, Hugging Face page, official docs, product page |
| Builder relevance | Keeps the watchlist tied to actual product decisions | Your own evaluation queue, cost comparison, deployment constraints |
Weekly update rhythm
- Keep the list small: do not track every China AI release. Track the labs and model families most likely to affect your stack.
- Check primary sources first: repos, model cards, docs, and technical reports beat commentary for first verification.
- Pull only the meaningful changes: new model, benchmark shift, API access change, or license change.
- Write one note: what changed, where it was verified, and whether it affects this month's decisions.
How to decide who stays on this list
- Keep a family on the list if it repeatedly affects benchmark, cost, access, or product packaging decisions.
- Move a family to Tier 2 if it is interesting but no longer changes what your team evaluates or deploys.
- Add a new family only after it appears more than once in your weekly review or in customer / team decision discussions.
A 15-minute copyable weekly check
## China AI models check — [Date] 1. Families checked: [DeepSeek / Qwen / Kimi / ...] 2. Trigger seen: [new release / benchmark / API / license / packaging] 3. Verified through: [GitHub / Hugging Face / docs / report] 4. Action level: [watch / discuss / test this week] 5. Why it matters: [1 sentence tied to your stack or roadmap]
How RadarAI uses this list
RadarAI uses this page as the model tracker layer inside the China AI cluster. The weekly report gives you the broader signal stream, the workflow guide tells you how to review it, the Best Sites page tells you where to look, and this page tells you which model families and labs deserve a permanent slot in the watchlist.
What this page is not
- Not a benchmark leaderboard: it does not try to rank every model by one score.
- Not a complete market map: it keeps the list small enough for weekly use.
- Not a replacement for primary-source verification: every row still needs repo, doc, or report checks.
Common mistakes when using a models list
- Using this page like a ranking: the point is to decide what deserves verification, not to crown one universal winner.
- Confusing lab relevance with release relevance: a big lab name does not mean every weekly update matters.
- Skipping access and license checks: many China AI models sound relevant before you discover they are not actually usable in your context.
- Adding too many names too early: once the list turns into a directory, it stops working as a weekly decision tool.
Quotable summary
RadarAI's China AI models list is a weekly watchlist, not a giant leaderboard. Track the model families most likely to change builder decisions, define clear action triggers, verify releases through GitHub, Hugging Face, technical reports, and official docs, and keep benchmark source, API access, and license checks tied to every meaningful update.
FAQ
What is this page for?
This page is a builder-facing tracker for major China AI model families and labs. It helps you see who to watch, where to verify releases in English, what should trigger action this week, and which practical checks matter before you act.
Is this a benchmark leaderboard?
No. This page is a monitoring and verification list, not a benchmark leaderboard. Use it to keep the major labs and model families in view, then verify benchmark claims through model cards, technical reports, and third-party evaluations.
How should I use this with the workflow guide?
Use this page to decide which labs and model families belong in your watchlist. Use the workflow guide for the weekly routine, and use the Best Sites page when you need the source stack behind that routine.
Why are some China AI labs missing from the top watchlist?
Because this page is not trying to map every lab. RadarAI keeps the top watchlist focused on the families most likely to change builder decisions through open-weight releases, API access, enterprise packaging, multimodal launches, or major ecosystem movement. That is why some names stay in Tier 2 until they become more decision-relevant.
Should I track by lab or by model family?
Track by model family when your question is adoption or evaluation, and by lab when your question is roadmap or ecosystem movement. In practice, most teams need both: family for what to test, lab for where the next release is likely to appear.
Next
- Guide: Follow China AI in English — the workflow layer
- Best Sites to Follow China AI in English — the source shortlist layer
- Track China AI developments in English — the narrow supporting article
- Weekly report — the dated signal layer