Zhipu GLM-5.3-Flash Details Surface: 320B Parameters, 18B Active, Hybrid Attention Built to Activate Domestic Compute

A research report from Guolian Minsheng Securities' computer team (吕伟/胡又文团队), republished via Sina Finance, describes how Zhipu's next-generation GLM-5.3-Flash model is being adapted at the architecture and system level for domestic Chinese compute — a telling sample of how Chinese frontier models are being built around homegrown silicon.

According to the report, GLM-5.3-Flash has roughly 320B total parameters, similar to GLM-4.5's 355B, but its active parameters drop from 32B to 18B and its layer count falls from 92 to 45.

Architecturally, the model pairs linear attention with sparse attention, introducing an IndexPool indexer for the sparse path. The design sharply cuts the memory and bandwidth pressure of attention computation during inference.

The report argues that sparse attention and low active parameters have become a primary direction for Chinese model optimization, with leading vendors' next-generation architectures converging on similar solutions.

It also notes that DeepSeek-V4 compresses context at the token level with DSA sparse attention, making 1M-token context a standard feature across its services — evidence of the same efficiency race among Chinese labs.

Shrinking active parameters lowers inference memory footprint and bandwidth demands, making frontier-class models far more practical on domestic accelerators. That is the substance behind the report's claim that GLM-5.3-Flash "fully activates" domestic compute.

What to watch next: real-world inference performance of GLM-5.3-Flash on domestic clusters, how the IndexPool sparse attention scales at long context, and the pace of follow-on releases in the GLM line.