Inside GLM-5.3-flash: beats GLM-5.2 at 1/10 cost, active params halved to 18B

baseten · x · 2026-08-27

A breakdown of GLM-5.3-flash's efficiency story:

Outperforms GLM-5.2 across domains at roughly 1/10 the cost.
Total params stay at GLM-4.5 scale, but active params dropped from 32B to 18B and layers from 92 to 45 — both nearly halved.
Hybrid linear + sparse attention cuts per-layer attention compute, and a smaller average KV cache compounds on longer sequences.
Visual intelligence plus RL: coding gives the model a proxy for expressing and testing world knowledge to improve multimodality.
The creation of flash was co-authored by the GLM-5.3 infrastructure agent, handling kernels, finding performance bottlenecks, and improving the serving stack.

More from Models