5 Chinese AI Models You Can Download Right Now (Open Weights)
Five open-weight Chinese LLMs you can download today — Qwen3.8-Flash-Next, MiniMax-M3, GLM-5.3, DeepSeek-V4-Pro and Kimi K3 — with the official Hugging Face links.

China's open-weight labs keep shipping, and five of their newest models are downloadable right now. Below is the list from the video, ranked from five to one, with the parameter count, the context window, the license note and the official Hugging Face card for each. Every detail here comes from the model's own card — no benchmark runs, no second-hand spec sheets — so where a claim belongs to the lab rather than an independent test, it is labelled that way.
The five open-weight models, ranked 5 to 1

Counting down from five, with a short note on what each one is actually useful for.
5. Qwen3.8-Flash-Next (Alibaba) — 180B parameters
Qwen3.8-Flash-Next is a 180 billion parameter preview of the architecture that will underpin Qwen4. The word "preview" is the important part: this is not the finished Qwen4, it is an early look at the design the next Qwen generation will be built on.
Why it matters: if you build on Qwen today, this card tells you where Alibaba's open models are heading next. Read the architecture notes now and you will not be surprised when Qwen4 lands. Because it is explicitly a preview, the sensible use is as a signal — a look at the shape of the next generation — rather than something to swap into a running pipeline the week it appears.

Official page: Qwen/Qwen3.8-Flash-Next
4. MiniMax-M3 — ~428B total, ~23B active, 1M context
MiniMax M3 has roughly 428 billion parameters in total, but only about 23 billion active at a time. That is a sparse mixture-of-experts setup: a large pool of experts with a small slice of them working per token.
Why it matters: you get the capacity of a very large model with the per-token cost of a much smaller one. The total tells you how much the model holds; the active figure tells you what a token roughly costs to run. MiniMax M3 also ships with a one million token context window, which is aimed at long inputs — full repositories, long transcripts, large documents in a single pass.

Official page: MiniMaxAI/MiniMax-M3
3. GLM-5.3 (Z.ai) — the coding pick
GLM 5.3 comes from Z.ai, and its makers describe it as the most capable open-weights model for coding. That is their claim on their own card, not an independent benchmark result — treat it as a strong signal of intent, not proof.
Why it matters: coding is the workload most developers actually judge a model on. If you already run an open model for code completion or agent work, GLM 5.3 is the one on this list worth testing against your own tasks and your own repo before you switch anything. A card-level claim about coding is best settled by the only benchmark that matters for you: your own codebase.

Official page: zai-org/GLM-5.3
2. DeepSeek-V4-Pro — 1.65T parameters, MIT license
DeepSeek V4 Pro is the official release of DeepSeek's V4 Pro model (build 0813) and it carries 1.65 trillion parameters. The headline, though, is the license: MIT.
Why it matters: MIT lets you use the weights for anything — commercial products, internal tools, fine-tunes you redistribute. On a list where licensing varies from model to model, MIT is the least friction you can get. If a permissive license is a hard requirement in your company, this is the one to look at first, and the one whose card you should read to see exactly what the release contains.

Official page: deepseek-ai/DeepSeek-V4-Pro-0813
1. Kimi K3 (Moonshot) — 2.8T params, 1M context
Kimi K3 from Moonshot takes the top spot: 2.8 trillion parameters, a one million token context window, and the biggest open model on this list so far.
Why it matters: this is the current ceiling for openly available models in this group. Two caveats before you plan anything. First, "biggest so far" is a moving target — the next release changes the ranking. Second, size is a practical constraint: read the card's requirements before you assume you can self-host it, and check how the context window behaves at your own input lengths.

Official page: moonshotai/Kimi-K3
What to check on each card before you download

Four things on these cards decide whether a model is actually usable for you, and all four sit plainly on the Hugging Face page.
Total parameters versus active parameters. MiniMax M3 reports both — roughly 428 billion in total, about 23 billion active — and the gap is the entire point of a sparse mixture-of-experts design. A card that gives you a single parameters number hides that distinction, so read closely when the two are reported separately.
The context window against your real inputs. One million tokens appears on both MiniMax M3 and Kimi K3, but the useful question is what your inputs look like, not what the ceiling is. If you want a refresher on what a token actually is, that is the unit every context window is measured in — and it is what turns a headline figure into a length of code, transcript or document.
The license. DeepSeek V4 Pro's card says MIT. On this list licensing varies from model to model, so read the license section on each card rather than assuming the permissive one applies across the board.
Whether the claim is the lab's or a benchmark's. GLM 5.3's coding line is Z.ai's own description of its own model. That does not make it wrong; it makes it a claim to test against your workloads. And if you are weighing open weights against a hosted model for automation work, it is worth making that comparison explicitly before you pull a multi-hundred-gigabyte checkpoint — when you do switch models mid-pipeline, structured output and retry handling are the parts to re-test first.
Quick recap

- #5 Qwen3.8-Flash-Next — 180B, preview of the architecture that will underpin Qwen4.
- #4 MiniMax-M3 — ~428B total, ~23B active, 1M token context.
- #3 GLM-5.3 — its makers call it the most capable open-weights model for coding.
- #2 DeepSeek-V4-Pro — 1.65T parameters, MIT license.
- #1 Kimi K3 — 2.8T parameters, 1M context, biggest open model so far.
- All five are open weights on Hugging Face, and every number above comes from the model's own card rather than a benchmark run.
Get the cheat sheet (PDF)
5 Chinese AI Models You Can Download Right Now
All 6 parts in this series
- Part 015 Chinese AI Models You Can Download Right Now (Open Weights)
- Part 02RemoveMacAI: remove Apple Intelligence on macOS 27
- Part 03LLMs vs AI agents vs workflows: which one are you building?
- Part 04Apex Governor Limits: Why Query 101 Kills Your Transaction
- Part 05What Is a Token in an LLM? Tokens vs Words, Explained
- Part 06Calling an LLM From Apex: Where Python Still Wins
AI Automation Engineer & Salesforce Developer
Salesforce developer since June 2018. I build AI agents, n8n workflows and Salesforce integrations, and write up what holds up in real use.



Comments
Loading comments...