GLM-5.3-Flash 320B-A18B — VRAM, speed & local setup
Written by Jakub Rusinowski · Last updated
320 billion parameters must all be resident — about 194 GB at Q4_K_M, so this is a multi-GPU or large-unified-memory model, not a single-card one. Only 18B activate per token, which is why it decodes at roughly the speed of a 18B dense model once loaded. Natively multimodal across text, image and video with a 1,048,576-token context. MIT licensed. Ollama publishes only a `:cloud` tag; local runners use the community GGUF conversions.
GLM-5.3-Flash 320B-A18B needs about 194 GB of VRAM at Q4_K_M — quantized weights plus framework overhead, before any KV cache. On Apple Silicon that figure comes out of unified memory.
VRAM and speed by quantization
Quoted against NVIDIA RTX 4090 (24 GB). Weights plus overhead only: this model's architecture is unpublished, so no KV cache is added.
| Quant | Memory | VRAM | Speed (est.) | Fit |
|---|---|---|---|---|
| Q2_K 2.63 bpw | 106 GB | — | Too big | |
| Q3_K_M 3.41 bpw | 137.2 GB | — | Too big | |
| Q4_K_M 4.83 bpw | 194 GB | — | Too big | |
| Q5_K_M 5.67 bpw | 227.6 GB | — | Too big | |
| Q6_K 6.56 bpw | 263.2 GB | — | Too big | |
| Q8_0 8.50 bpw | 340.8 GB | — | Too big | |
| F16 16.00 bpw | 640.8 GB | — | Too big |
Black marker = usable memory on the NVIDIA RTX 4090 (24 GB). Estimates from the memory-bandwidth roofline on the methodology page. GLM-5.3-Flash 320B-A18B VRAM calculator →
Get GLM-5.3-Flash 320B-A18B running
The cheapest catalogued GPU that runs GLM-5.3-Flash 320B-A18B is the Apple M3 Ultra (512 GB).
As an Amazon Associate we earn from qualifying purchases. Cloud GPU links are referral links — we may earn a commission at no extra cost to you.
How to run GLM-5.3-Flash 320B-A18B
The Ollama tag `glm-5.3-flash:cloud` is cloud-hosted, so there is no local run command for this model.
Specifications
Verified — Checked against the primary source — the model card or the vendor spec page — and corroborated by a second independent source.
- Parameters
- 320 Billion (18B active)
- Context window
- 1,048,576
- Architecture
- Hybrid-Attention MoE (multimodal)
- Provider
- Zhipu AI (Z.ai)
- Licence
- MIT
- Practical deployment
- Workstation (large unified memory or pro GPU)
- Specified at
- Q4_K_M
- System RAM
- 256 GB
- Record updated
- 2026-09-11
Commercial use permitted. No usage restrictions beyond attribution.
See also GLM-5.3
Quality and use cases
Scores as published by the model’s authors or an independent evaluator — quality, not throughput, and not measured by us.
| Benchmark | Score | Provenance |
|---|---|---|
| DeepSWE | 63.4 / 100 % | vendor-claimed · https://www.marktechpost.com/2026/08/26/z-ai-releases-glm-5-3-flash-a-320b-a18b-natively-multimodal-moe-with-a-1m-token-context/ |
Independent evaluation — not vendor-claimed and not measured by us
| Benchmark | Score | Provenance |
|---|---|---|
| AA Intelligence Index v4.3.2 | 42 | independent · Artificial Analysis · snapshot 2026-10-13 |
Can I run GLM-5.3-Flash 320B-A18B on my GPU?
- GLM-5.3-Flash on AMD Ryzen AI Max+ 395
- GLM-5.3-Flash on Apple M1 Ultra
- GLM-5.3-Flash on Apple M2 Max
- GLM-5.3-Flash on Apple M2 Ultra
- GLM-5.3-Flash on Apple M4 Max
- GLM-5.3-Flash on Apple M5 Max
- GLM-5.3-Flash on NVIDIA A100 80GB (PCIe)
- GLM-5.3-Flash on NVIDIA DGX Spark
- GLM-5.3-Flash on NVIDIA H100 80GB (PCIe)
- GLM-5.3-Flash on NVIDIA RTX PRO 6000 Blackwell
GLM-5.3-Flash 320B-A18B — frequently asked questions
How much VRAM does GLM-5.3-Flash 320B-A18B need?
About 194 GB at Q4_K_M — quantized weights plus framework overhead, before any KV cache. The cache grows with context length and is added on top; the table above folds it in. Apple Silicon counts unified memory toward the same figure.
Does GLM-5.3-Flash 320B-A18B run on an RTX 4090 (24 GB)?
No. GLM-5.3-Flash 320B-A18B needs about 194 GB at Q4_K_M, more than a single RTX 4090's 24 GB. It needs a larger card, several GPUs, or Apple Silicon with enough unified memory — or it runs with part of the weights offloaded to system RAM, which is much slower.
How do I run GLM-5.3-Flash 320B-A18B locally?
The Ollama tag `glm-5.3-flash:cloud` is cloud-hosted, so there is no local run command for this model. Running the published tag would send your prompts to a hosted GPU rather than your own machine.