The capabilities, trade-offs and people behind China’s AI models.
Tencent Hunyuan released the Hy Translation app on September 24, 2026, bringing its open-source Hy-MT2 translation model to consumer devices with offline support, 33-language coverage including five Chinese minority languages, and availability across 12 countries and regions.
SenseTime is moving U1 Pro into production use, with enterprise API access handled through a whitelist.
Xiaomi released Pro and Flash variants of MiMo-V2.6, alongside a smaller distilled model and research materials.
Third-party developer Unsloth has published GGUF and FP8 quantizations of Alibaba's open-weight Qwen-Image-2.1 model, enabling local inference on consumer hardware with as little as 6GB VRAM via offloading. However, Unsloth's own documentation cautions that memory figures are estimates and recommends FP8 over GGUF for GPU users, while the underlying Qwen Research License Agreement restricts all use to non-commercial research and evaluation.
Qwen released downloadable weights for an image model that supports transparent output and editing with multiple reference images.
MiniMax Code's published repository covers terminal, headless and editor-integration components, while its desktop application remains separate.
