China’s open LLM ecosystem grows more selective at the frontier

By mid-2025, China had become a leading force in open-source LLMs, with state media reporting that the country accounted for 1,509 of the world’s ~3,755 publicly released LLMs. The ecosystem spans major technology companies, startups and academic labs, with models distributed through Hugging Face, GitHub, cloud APIs and domestic repositories.

Alibaba’s Qwen series, DeepSeek’s efficient MoE models, Zhipu AI’s ChatGLM and GLM releases, Moonshot AI’s Kimi line, Baichuan AI, MiniMax and Fudan University’s MOSS are among the most prominent examples. Releases cover general-purpose chat, coding, reasoning, multimodal input and domain-specific applications in areas such as law, healthcare, education and finance.

Through early July 2026, release activity accelerated toward coding and agentic systems. DeepSeek moved to V4, Alibaba shipped Qwen3.5 and Qwen3.6 open-weight families, Z.ai advanced GLM models, Moonshot released Kimi K2.7-Code, MiniMax pushed its M-series toward agents, and Tencent opened Hy3 under Apache 2.0. At the same time, newer flagship systems such as Qwen3.7-Plus and Baidu’s Ernie 5.0 and Ernie 5.1 remained API-only, showing a shift toward opening mid-tier and prior-generation models while holding back some frontier weights.