Search
The new generation of large models is shifting towards RSI self-recursive upgrades, and Zhipu has announced the latest progress: its recently released multimodal large model, GLM-5.3 Flash, is deployed using 100,000 domestically-produced chips, with a significant 80% reduction in unit Token inference costs compared to the beginning of the year. Driven by GLM-5.3, the Infra Agent enabled the model to run smoothly on domestic accelerator cards in just two weeks, handling all online traffic. The end-to-end throughput increased to 3.2 times the original, while addressing three key issues, including KDA accuracy drift. Additionally, Zhipu revealed that the next-generation large model, GLM-6, will achieve fully self-training and plans to focus on researching the model's autonomous judgment stopping and self-correction capabilities.
-
C114 Communication Network -
Communication Home
