High-memory GPUs on OneInfer
GLM-5.3-Flash at BF16 needs ~328 GB of VRAM; FP8 halves that. These are current OneInfer listings by VRAM, for scoping a self-host deployment.
VRAM by precision
| Precision | Catalog-reported footprint | Notes |
|---|---|---|
| BF16 | 328 GB model size / 350 GB min VRAM | Catalog-reported, 29 Aug 2026. Multi-GPU datacenter deployment only. BF16 is the published weight format. |
| FP8 | ~164 GB (estimated) | Halves memory at typical accuracy cost for this model class. Verify against an FP8 checkpoint release before deploying. |
| INT4 | ~82 GB (estimated) | Fits a single H100 (80 GB) only with KV-cache offload and aggressive quantization. Expect quality caveats on long-context reasoning. |
How these figures are calculated
BF16 row is sourced directly from the OneInfer catalog entry (`model_size_gb = 328`, `model_min_vram = 350`). FP8 and INT4 rows are derived by halving and quartering the BF16 weight size respectively — they are estimates, not catalog-reported numbers. Multiply by ~1.2 to budget for sustained inference at the 1M-token context (KV cache + activations).
Ready to test the workflow?
Create account & add creditsArchitecture
320B-parameter MoE (18B active per token), BF16 weights, 1,048,576-token native context, native multimodal input (text + image), text output, tool calling supported. Source: OneInfer catalog entry for `z-ai/GLM-5.3-Flash`, 29 Aug 2026.
API is the default deployment story
GLM-5.3-Flash weights are not yet released as of this page's last verification. Treat the BF16 row above as the minimum spec for an eventual self-host deployment and use the API route on OneInfer for production traffic in the meantime.
Frequently asked questions
What GPU do I need to run GLM-5.3-Flash?
BF16 weights need ~328 GB of VRAM plus headroom — at minimum a multi-GPU datacenter deployment (catalog-reported min VRAM is 350 GB). FP8 halves that to ~164 GB; INT4 fits a single H100 with KV-cache offload at ~82 GB.
Is GLM-5.3-Flash open source?
Weights were not released as of this page's last verification. Treat every figure above as an architecture-driven estimate and use the API route on OneInfer for production traffic.
How is VRAM estimated?
The BF16 row is catalog-reported (`model_size_gb = 328`, `model_min_vram = 350`). FP8 and INT4 rows are derived by halving and quartering the BF16 weight size — verify against the Hugging Face file listing when weights ship.
Where is the live GLM-5.3-Flash model page?
The canonical model page with current OneInfer pricing, capabilities, and availability is /models/zai-org/GLM-5.3-Flash. This page is a focused facet of that entity, not a replacement for it.
How should I treat benchmark or price claims?
Check each claim's provenance label and observed date. Vendor-reported and independently verified numbers are shown as separate evidence classes on this hub.
Put GLM-5.3-Flash to work
Fund a controlled evaluation, start with a prepared prompt, and measure quality and cost on your own workload.