GLM - Miles

Miles ships RL recipes for every GLM generation currently in production: the dense GLM4 line (9 B, 32 B — Zhipu “Z1” reasoning checkpoints), the GLM4.5 MoE at 106 B-A12B and 355 B-A32B, the compact GLM4.7 Flash with 64 routed experts, and the 744 B-A40B GLM5 flagship.

Variants

Family	Class	Sizes	Recipe
GLM4	Dense	9 B · 32 B	glm4
GLM4.5	MoE	12 B / 106 B · 32 B / 355 B	glm4-5
GLM4.7 Flash	MoE (64 experts, top-4)	Compact	glm4-7-flash
GLM5	MoE	40 B / 744 B	glm5

Fastest path to train

GLM4-9B (GLM-Z1-9B-0414) on a single 8× H100 node — the smallest GLM recipe:

cd /root/miles
hf download zai-org/GLM-Z1-9B-0414 --local-dir /root/GLM-Z1-9B-0414
bash scripts/run-glm4-9B.sh

See the GLM4 Dense page for weight conversion and the full walkthrough.

Which variant do I pick?

Single-node GLM first try → GLM4-9B (glm4).
Larger dense → GLM4-32B (glm4).
MoE on a budget → GLM4.5-106B-A12B (glm4-5).
Full MoE scale (multi-node) → GLM4.5-355B-A32B (glm4-5).
Compact MoE for routing experiments (R3) → GLM4.7 Flash (glm4-7-flash).
Frontier scale (744 B) → GLM5 (glm5).

Qwen3 MoE GLM-5 / GLM-5.1