LLM Model Notes
Model tracker
Open-weight models from Chinese labs
Recent open-weight releases, newest repository first within each lab. Every value comes from the model's own repository: the model card, its configuration file, or its license file. A blank or general entry means the card does not state more.
Checked . Dates are the month the repository was created, which can differ from the announcement. Read how to read a model card before comparing rows.
| Lab | Model | Repository created | Architecture as stated | License |
|---|---|---|---|---|
| DeepSeek | DeepSeek-V4.1-Flash | Multimodal MoE, context up to 1M | MIT | |
| DeepSeek | DeepSeek-V4-Pro | MoE, 1.6T total / 49B active | MIT | |
| DeepSeek | DeepSeek-V4-Flash | MoE, 284B total / 13B active | MIT | |
| Alibaba Qwen | Qwen3.8-2.4T-A95B | MoE, 2.4T total / 95B active, 262K native | Qwen3.8-Max License | |
| Alibaba Qwen | Qwen3.8-27B | 27B, 262K native | Apache-2.0 | |
| Alibaba Qwen | Qwen3.6-35B-A3B | MoE, 35B total / 3B active | Apache-2.0 | |
| Moonshot AI | Kimi K3 | MoE, 2.8T total / 104B active, 1M context | Kimi K3 License | |
| Moonshot AI | Kimi-K2.6 | MoE, 1T total / 32B active | Modified MIT | |
| Z.ai | GLM-5.3 | MoE, 1M context | GLM-5.3 License | |
| Z.ai | GLM-5.2 | MoE, 1M context | MIT | |
| MiniMax | MiniMax-M3.1-Flash | Multimodal MoE, about 428B total / 23B active | MiniMax Community License | |
| MiniMax | MiniMax-M2 | MoE, 230B total / 10B active | Modified MIT | |
| Tencent | Hy3 | MoE, 295B total / 21B active | Apache-2.0 | |
| StepFun | Step 5 Preview | Sparse MoE, 600B total / 27B active | StepFun Model License 1.0 | |
| Xiaomi MiMo | MiMo-V2.6-Pro-RL | Sparse MoE, 1.02T total / 42B active | MIT | |
| ByteDance Seed | Seed-OSS-36B-Instruct | 36B, 512K context | Apache-2.0 | |
| Baidu | ERNIE-4.5-300B-A47B-PT | MoE, 300B total / 47B active | Apache-2.0 |
How rows are chosen
The tracker lists general-purpose language models with public weights. It leaves out hosted-only products, image and speech models, and quantized copies. A row is added only after its facts are read on the repository; announcements and third-party summaries are not used as sources.
Licenses differ inside one model family, so each row names the license of that repository. A custom license name means the terms are specific to the release and need to be read in full.
For releases that are not yet checked here, see the automatically refreshed news page.