howtospark
models / glm-5-2

GLM-5.2

zai-org/GLM-5.2
params
753B
active params
51.0B
sparsity
MoE
attention
Full attention
architecture
glm_moe_dsa
license
mit
downloads
392.7k
likes
3.8k
hf updated
Jul 2, 2026

Visualize

total size

each column is a layer — lit blocks are the experts that fire per token

bits / param
16-bit
footprint
1.40 TiB
sparks to fit
13×
block ≈
5.89B params
active paramsdormant expertsghost = precision traded awayblock fill ∝ bits kept · GLM-5.2753.32994048B params

Benchmarks (0)

No benchmarks for this model yet.