Z.ai released GLM-5.3, a coding and cybersecurity-focused model built on post-training of its 743B base model rather than new pretraining. The release includes specific benchmark scores such as Terminal Bench 3.0 at 28.3 and DeepSWE at 66.9.
Alibaba released Qwen3.8-27B, a native multimodal dense model with 262K context extendable to 1M via YaRN, positioned for local use on 17GB RAM. Meanwhile, Cursor announced it is joining SpaceXAI to work across Grok and related products.
The article highlights a trend of Chinese labs specializing in open-weight models and agent runtimes becoming optimization targets, alongside skepticism toward vendor benchmark claims.