High-memory GPUs on OneInfer
A single-GPU instance below is rarely enough for a model this size — expect a multi-GPU node. These are current listings by VRAM, for scoping a self-host deployment once weights are available.
VRAM by precision
| Precision | Approximate VRAM | Notes |
|---|---|---|
| FP16 / full precision | ~1.5 TB | Multi-GPU datacenter deployment only. |
| FP8 | ~750 GB | Halves memory at typical accuracy cost for this model class. |
| 2-bit quantized | ~245 GB | The realistic self-host floor; expect measurable quality loss. |
Weights are not released yet
Self-hosting GLM-5.3 requires open weights, which Z.ai has not shipped as of this page’s last verification. Treat every figure above as an architecture-driven estimate until a real checkpoint and license are confirmed.
Ready to test the workflow?
Architecture
GLM-5.3 uses a mixture-of-experts design: roughly 744B total parameters with about 40B active per token, which is why its serving cost sits well below a dense model of comparable total size.