GLM-4.7-Flash · vLLM · FP8

status: blocked

Configuration

Modelzai-org/GLM-4.7-Flash
CompanyZhipu AI
FamilyGLM
Parameters31B (MoE)
EnginevLLM
Quant / precisionFP8
Why this quantNear-BF16 quality at half the bytes; official FP8 weights published.
Downloadzai-org/GLM-4.7-Flash
Context window131072
Input modalitiestext

Measured results

Prefill tok/s
Decode tok/s
Peak memory (GB)

Full run command

# blocked — no confirmable repo (see Notes)

glm-4.7-flash Zhipu AI GLM FP8 16-40B conc-32

Notes

Blocked — the named model does not resolve to a real HF repo, and no clean substitute fits the intended slot. Verified 2026-06-22.

  • zai-org/GLM-4.7-Flash (and glm-4-7-flash, GLM-4-Flash) → 404 on HF. The stub describes a ~31B MoE GLM “Flash”, which does not exist yet.
  • The nearby real GLMs don’t match: zai-org/GLM-4.5-Air is a 106B/12B MoE (belongs in the 41-130B / 130B+ buckets, not 16-40B), and zai-org/GLM-4-32B-0414 is a 32B dense model, not a Flash MoE. Substituting either would change the model class this config is meant to capture.
  • Per the “when unsure, BLOCK — don’t guess” policy, left for human review rather than silently swapped. If the intent is the dense 32B, point source_repo at zai-org/GLM-4-32B-0414; if it’s the small MoE, use zai-org/GLM-4.5-Air (and re-bucket by size).