← Back to live feed1 story
Zai deployed GLM-5.3 and GLM-5.3-Flash models to agent platforms including Devin CLI, Droid, and the FLock API. The open-weight GLM-5.3 model achieved a 50% gain in coding performance over GLM-5.2 on the Code Bench, with specific improvements in complex programming and long-horizon agent tasks.
The GLM-5.3-Flash version focuses on efficiency using a 30T-token multimodal training corpus and a combination of sparse and linear attention. In the Agent Arena, the Flash model ranks 4th among open source models and 19th overall based on 9,000 real-world sessions, recording a $0.12 median cost per task.
Sign in to suggest edits
Key sources
- SOURCE@arena“At a $0.12 median cost per task and a +4.6% net improvement, it has reshaped the Pareto frontier!”x.com
- SUPPORT@dabit3“rolled out support for GLM 5.3 Flash along with an improved model picker UX in Devin CLI”x.com
- SUPPORT@factoryai“GLM-5.3 and GLM-5.3 Flash have arrived in Droid”x.com
- SUPPORT@factoryai“stronger results on complex programming and long-horizon tasks”x.com
- SUPPORT@flock_io“combining sparse + linear attention, mHC, and a 30T-token multimodal training corpus”x.com
- SUPPORT@deeplearningai“This unexpected jump in exploit generation prompted a temporary safety hold on the model weights’ release”x.com