5 Chinese AI Models You Can Download Right Now (Open Weights)

Five Chinese labs currently ship open-weight models you can download yourself, and this list runs from number five to number one. Every entry below is on Hugging Face, with the parameter counts, context windows and license terms taken from that model's own official card. These are the same five models and the same links shown in the video.

5. Qwen 3.8 Flash Next (Alibaba)

Qwen 3.8 Flash Next is a 180 billion parameter model from Alibaba, and its card describes it as a preview of the architecture that will underpin Qwen 4. That framing matters: you are not looking at a finished generation, you are looking at the shape of the next one. It is also the smallest model on this list by parameter count, which makes it the most practical starting point if you just want to see where the Qwen line is going next.

Official model card: Qwen/Qwen3.8-Flash-Next

https://huggingface.co/Qwen/Qwen3.8-Flash-Next

4. MiniMax M3

MiniMax M3 lists roughly 428 billion total parameters, but only about 23 billion activated parameters — the Mixture-of-Experts pattern, where the full model holds the weights and a small slice runs per token. The card also lists a one million token context window, so long documents, large codebases and multi-file context fit in a single pass. If you work with long inputs, this is the entry on the list to read first.

Official model card: MiniMaxAI/MiniMax-M3

https://huggingface.co/MiniMaxAI/MiniMax-M3

3. GLM 5.3 (Z.ai)

GLM 5.3 comes from Z.ai, and its makers call it the most capable open-weights model for coding. Treat that as their claim rather than a settled result — open the card, read the reported evaluations, and run your own tasks against it before you rearrange your setup. Still, for developers who spend the day in an editor, a coding-focused open-weights model is the most directly useful thing on this list.

Official model card: zai-org/GLM-5.3

https://huggingface.co/zai-org/GLM-5.3

2. DeepSeek V4 Pro

DeepSeek V4 Pro is listed at 1.65 trillion parameters, and it ships under the MIT license. That license is the headline for anyone building a product: MIT is permissive, so you can use the model for essentially anything, including commercial work, without the extra terms some open-weight releases carry. If licensing is what usually blocks you from shipping an open model, this is the one to start with.

Official model card: deepseek-ai/DeepSeek-V4-Pro-0813

https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813

1. Kimi K3 (Moonshot)

Kimi K3 from Moonshot takes the top spot: 2.8 trillion parameters and a one million token context window, which makes it the biggest open model so far. Scale alone is not the reason to pick it — the combination of that size with a million-token context is. It is also the heaviest download here, so treat it as the one you plan infrastructure for rather than the one you pull on a whim.

Official model card: moonshotai/Kimi-K3

https://huggingface.co/moonshotai/Kimi-K3

Read the numbers before you download

Three details on these cards decide whether a model is usable for your setup. Total versus active parameters: MiniMax M3 lists about 428B total but roughly 23B active, which is a very different resource story from a dense model of the same headline size. Context window: MiniMax M3 and Kimi K3 both list one million tokens, which changes what you can feed in one call. License: DeepSeek V4 Pro is MIT, so check the terms on the others against your own use case before you build on them. And since these are specific model cards with their own release dates, re-read the repo page before you commit — cards get updated.

Quick recap

• #5 Qwen 3.8 Flash Next — 180B parameters, preview of the architecture behind Qwen 4
• #4 MiniMax M3 — ~428B total, ~23B active, 1M-token context window
• #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, usable for anything
• #1 Kimi K3 — 2.8T parameters, 1M-token context, biggest open model so far
• All five are open weights on Hugging Face — the links above are the official cards