Zhipu AI operates new GLM-5.3-Flash model on 100,000 Chinese chips

Zhipu AI, a Beijing-based company, has revealed that its new AI model, GLM-5.3-Flash (previously known by the codename Ox Alpha), was operated using more than 100,000 Chinese AI chips. The model gained significant traction on the Openrouter marketplace, reportedly exceeding 50 trillion tokens in traffic. According to benchmarks from Artificial Analysis, the model reached the top 10 of the Intelligence Index, performing similarly to Anthropic’s Claude Opus 4.

Following the release of the model, the US startup Abliteration.ai has developed a modified version titled ‘abliterated-model-large-v2’. This process, known as ‘abliteration’, involves modifying the model's weights to suppress internal activation patterns associated with refusal mechanisms. The goal is to create a version that is less likely to decline security-related or sensitive queries while maintaining coding and cyber capabilities. Abliteration.ai claims their modified GLM-5.3 version performs strongly on benchmarks like CyberGym and ExploitGym, though they note that direct comparisons to other models may be limited by varying testing conditions.