China
Open models from China
We track 195 open-weight models from 12 labs in China. They account for 78 of 205 tracked releases (38%) from Dec 2025 to Nov 2026. The newest is Qwen3.8-27B-DSpine-b16, listed on 7 Nov 2026.
Explore the models
Latest models
Newest listed firstCheapest tracked APIs
- DeepSeek-V4-FlashBaidu · FP8$0.0314 (₹2.96)
- Qwen3.6-35B-A3BPrimeIntellect · FP8$0.0465 (₹4.38)
- GLM-5.3-FlashRelace · FP4$0.0495 (₹4.67)
- Qwen3.5-9BPrimeIntellect · BF16$0.058 (₹5.47)
- DeepSeek-V4.1-Flash-1017Sail Research · FP4 (other providers mostly serve FP8)$0.0815 (₹7.69)
- DeepSeek-V4-Flash-0731Relace · FP4 (other providers mostly serve FP8)$0.0875 (₹8.25)
- DeepSeek-V4.1-FlashBaidu · FP8$0.0887 (₹8.37)
- Qwen3.8-27BPrimeIntellect · FP8$0.1218 (₹11)
Per million tokens, blended three parts input to one part output. INR at ₹94.29 per dollar.
How the work developed
Alibaba’s Qwen team publishes both base models and instruction-tuned checkpoints. Its Qwen2.5 release documented multilingual support and separate model sizes, giving developers choices for local deployment and fine-tuning. The model card links the weights to the team’s technical report and inference examples.
DeepSeek took a different research direction with its R1 work. The team described reinforcement learning for reasoning, released the main model’s weights under MIT, and published smaller distilled models based on Qwen and Llama. Those distilled checkpoints retain conditions from their underlying model families. China’s entries here cover the releasing labs; they do not imply that training data or contributors came from only one country.
Sources and counting method
Counts cover selected repositories, including model variants. Dates come from Hugging Face repository creation dates and may differ from public announcements. Location follows the lab record, not the nationality of individual contributors.