Language Models

DeepSeek, Alibaba's Qwen, Moonshot's Kimi, Zhipu's GLM and MiniMax made Chinese open-weight models a global default, and the ecosystem beneath them is broad: vertical models for power utilities, finance and healthcare; inference systems built for domestic GPUs and neuromorphic chips; training-efficiency research; and a growing number of on-device models for laptops and phones. Chip export controls have pushed much of this work toward efficiency and toward hardware from Huawei, Cambricon and other Chinese vendors.

Entries under language models cover releases that most English coverage skips: industry-specific models from state-owned enterprises, hybrid inference systems from carriers such as China Mobile, speed records on open training benchmarks, small-large collaborative inference frameworks and on-device models shipped with hardware partners. Frontier releases from the big labs appear here only when we add something the English press missed.