GLM-4.5 (355B-A32B) VRAM Requirements, KV Cache & Best GPUs
GLM-4.5 (355B-A32B) (355B total, 32B active, Sparse MoE + GQA) uses 200.75 GiB for 4-bit weights (Q4_K_M), requiring 204.19 GiB at 8K context and 212.81 GiB at 32K. · Hugging Face: zai-org/GLM-4.5 ↗ · GGUF: unsloth/GLM-4.5-GGUF ↗
Fits with Safe Headroom
Config-derived formula
HF config.json ↗
Required GPU Memory
204.19 GiB
219.25 GB (decimal)
Usable Hardware Memory
243.2 GiB
256 GB (physical)
Safe Headroom
+39.01 GiB
80% utilization
Est. Decode Throughput
~115.1 tok/s
32B active params
Est. Prefill Throughput
~10619 tok/s
4K Prompt TTFT: ~386 ms
Est. Cost / 1M Tokens
$9.17
Sparse MoE + GQA · MIT
Recommended Hardware for Current Configuration
Best Single GPU
Instinct MI325X 256GB (256GB)
~$21,000
Best Value / Used GPU
2x H200 141GB SXM
282GB · ~$64,000
Apple Unified Memory
Mac Studio M3 Ultra 512GB
~$9,500
Cloud From
DigitalOcean · $3.80 / hr
1x Instinct MI325X 256GB
Production Launch Commands
llama.cpp (GGUF)
llama-server -hf unsloth/GLM-4.5-GGUF:Q4_K_M \ --ctx-size 8192 -ngl 999 --flash-attn
vLLM
vllm serve zai-org/GLM-4.5 \ --max-model-len 8192 --gpu-memory-utilization 0.92
Live Token Speed Simulator (TTFT ~386ms · ~115.1 tok/s)
[GLM-4.5 (355B-A32B) · 204.19 GiB] allocated on target GPU. Prefill completed in TTFT window; streaming autoregressive tokens at memory-bandwidth-bound rate...
Cloud GPU Comparison & Rent-vs-Buy TCO
Verified public rates across RunPod, Vast.ai, Vultr, and DigitalOcean.
Purchase break-even
~25.5 months
Est. local power: $21.28 / mo
| Provider | Recommended Instance | Total VRAM | Hourly Rate | Monthly at Selected Hours | Billing Mode | Deploy |
|---|---|---|---|---|---|---|
| RunPod | 1x B300 288GB Blackwell Ultra (Community Cloud) | 288 GB | $6.94 / hr | $843.90 / mo | Per-second | Deploy GPU ↗ |
| Vast.ai | 2x H200 141GB SXM (Verified Host · NVL) | 282 GB | $7.74 / hr | $940.94 / mo | Spot / On-demand | Deploy GPU ↗ |
| Vultr ($300 Free Credit*) | 8x A100 80GB SXM (Dedicated Instance · $300 Credit*) | 640 GB | $19.18 / hr | $2331.80 / mo | Hourly | Deploy GPU ↗ |
| Vultr Cloud GPU | 8x A100 80GB SXM (Dedicated Instance) | 640 GB | $19.18 / hr | $2331.80 / mo | Hourly | Deploy GPU ↗ |
| DigitalOcean | 1x Instinct MI325X 256GB (Dedicated Instance) | 256 GB | $3.80 / hr | $462.08 / mo | Per-second | Deploy GPU ↗ |
Rates reflect public listings on the verification date and fluctuate with regional inventory.
Fine-Tuning VRAM Requirements (QLoRA / LoRA / Full) — GLM-4.5 (355B-A32B)
QLoRA (4b)
Fits with Safe Headroom
206.34 GiB
Model Weights: 200.75 GiB · Gradients + Optimizer: 4.48 GiB · Act: 1.11 GiB
Recommended GPU: Instinct MI325X 256GB (256GB)
LoRA (16b)
Exceeds Available GPU Memory
669.76 GiB
Model Weights: 661.24 GiB · Gradients + Optimizer: 7.04 GiB · Act: 1.48 GiB
Recommended GPU: 8x H200 141GB SXM (1128GB)
Full BF16
Exceeds Available GPU Memory
2713.68 GiB
Model Weights: 661.24 GiB · Gradients + Optimizer: 2049.84 GiB · Act: 2.6 GiB
Recommended GPU: Multi-GPU Required
Engineering Sizing Notes & Optimization Playbook
Quick Hardware Verdict
Instinct MI325X 256GB (256GB physical, 243.2 GiB usable) runs 4-bit (Q4_K_M) GLM-4.5 (355B-A32B) at 8K context using 204.19 GiB with 39.01 GiB of safe headroom (~115.1 tok/s single-stream decode).
KV Cache & Runtime Tuning
GLM-4.5 (355B-A32B) uses Sparse MoE + GQA (355B total, 32B active per token). Moving from 8K to 32K context expands FP16 KV cache from 2.88 GiB to 11.5 GiB; enabling 8-bit KV cache (llama.cpp -ctk q8_0 -ctv q8_0 or vLLM --kv-cache-dtype fp8) saves 5.75 GiB at 32K (reducing total VRAM to 207.06 GiB).
VRAM Scaling Across Context Windows (4K to 128K, 4-bit Weights)
GLM-4.5 (355B-A32B) × 49 GPUs (4-bit Q4_K_M)
| Target Hardware | Total VRAM | Memory Bandwidth | 4-bit @ 8K | 4-bit @ 32K | Decode Speed | Compatibility | Full Compatibility Report |
|---|---|---|---|---|---|---|---|
| NVIDIA RTX 3060 12GB | 12 GB (11.4 GiB) | 360 GB/s | 204.19 GiB | 212.81 GiB | ~2.8 tok/s | Exceeds Available GPU Memory | Full Compatibility Report → |
| NVIDIA RTX 4060 Ti 16GB | 16 GB (15.2 GiB) | 288 GB/s | 204.19 GiB | 212.81 GiB | ~2.8 tok/s | Exceeds Available GPU Memory | Full Compatibility Report → |
| NVIDIA RTX 4070 Ti Super | 16 GB (15.2 GiB) | 672 GB/s | 204.19 GiB | 212.81 GiB | ~2.8 tok/s | Exceeds Available GPU Memory | Full Compatibility Report → |
| NVIDIA RTX 4080 Super | 16 GB (15.2 GiB) | 736 GB/s | 204.19 GiB | 212.81 GiB | ~2.8 tok/s | Exceeds Available GPU Memory | Full Compatibility Report → |
| NVIDIA RTX 5070 12GB | 12 GB (11.4 GiB) | 672 GB/s | 204.19 GiB | 212.81 GiB | ~2.8 tok/s | Exceeds Available GPU Memory | Full Compatibility Report → |
| NVIDIA RTX 5070 Ti | 16 GB (15.2 GiB) | 896 GB/s | 204.19 GiB | 212.81 GiB | ~2.9 tok/s | Exceeds Available GPU Memory | Full Compatibility Report → |
| NVIDIA RTX 5080 | 16 GB (15.2 GiB) | 960 GB/s | 204.19 GiB | 212.81 GiB | ~2.9 tok/s | Exceeds Available GPU Memory | Full Compatibility Report → |
| NVIDIA RTX 3090 | 24 GB (22.8 GiB) | 936 GB/s | 204.19 GiB | 212.81 GiB | ~3 tok/s | Exceeds Available GPU Memory | Full Compatibility Report → |
| NVIDIA RTX 4090 | 24 GB (22.8 GiB) | 1008 GB/s | 204.19 GiB | 212.81 GiB | ~3 tok/s | Exceeds Available GPU Memory | Full Compatibility Report → |
| NVIDIA RTX 4090 D | 24 GB (22.8 GiB) | 1008 GB/s | 204.19 GiB | 212.81 GiB | ~3 tok/s | Exceeds Available GPU Memory | Full Compatibility Report → |
| NVIDIA RTX 5090 | 32 GB (30.4 GiB) | 1792 GB/s | 204.19 GiB | 212.81 GiB | ~3.1 tok/s | Exceeds Available GPU Memory | Full Compatibility Report → |
| NVIDIA RTX 5090 D | 32 GB (30.4 GiB) | 1792 GB/s | 204.19 GiB | 212.81 GiB | ~3.1 tok/s | Exceeds Available GPU Memory | Full Compatibility Report → |
| NVIDIA RTX A6000 | 48 GB (45.6 GiB) | 768 GB/s | 204.19 GiB | 212.81 GiB | ~3.3 tok/s | Exceeds Available GPU Memory | Full Compatibility Report → |
| NVIDIA RTX 6000 Ada | 48 GB (45.6 GiB) | 960 GB/s | 204.19 GiB | 212.81 GiB | ~3.4 tok/s | Exceeds Available GPU Memory | Full Compatibility Report → |
| NVIDIA RTX PRO 6000 Blackwell | 96 GB (91.2 GiB) | 1792 GB/s | 204.19 GiB | 212.81 GiB | ~4.8 tok/s | Exceeds Available GPU Memory | Full Compatibility Report → |
| NVIDIA DGX Spark (GB10 128GB) | 128 GB (99.8 GiB) | 273 GB/s | 204.19 GiB | 212.81 GiB | ~4.5 tok/s | Exceeds Available GPU Memory | Full Compatibility Report → |
| NVIDIA L4 24GB | 24 GB (22.8 GiB) | 300 GB/s | 204.19 GiB | 212.81 GiB | ~2.9 tok/s | Exceeds Available GPU Memory | Full Compatibility Report → |
| NVIDIA A40 48GB | 48 GB (45.6 GiB) | 696 GB/s | 204.19 GiB | 212.81 GiB | ~3.3 tok/s | Exceeds Available GPU Memory | Full Compatibility Report → |
| NVIDIA L20 48GB | 48 GB (45.6 GiB) | 864 GB/s | 204.19 GiB | 212.81 GiB | ~3.3 tok/s | Exceeds Available GPU Memory | Full Compatibility Report → |
| NVIDIA L40S 48GB | 48 GB (45.6 GiB) | 864 GB/s | 204.19 GiB | 212.81 GiB | ~3.3 tok/s | Exceeds Available GPU Memory | Full Compatibility Report → |
| NVIDIA A800 80GB SXM | 80 GB (76 GiB) | 2039 GB/s | 204.19 GiB | 212.81 GiB | ~4.2 tok/s | Exceeds Available GPU Memory | Full Compatibility Report → |
| NVIDIA A100 80GB SXM | 80 GB (76 GiB) | 2039 GB/s | 204.19 GiB | 212.81 GiB | ~4.2 tok/s | Exceeds Available GPU Memory | Full Compatibility Report → |
| NVIDIA H800 80GB SXM | 80 GB (76 GiB) | 3350 GB/s | 204.19 GiB | 212.81 GiB | ~4.3 tok/s | Exceeds Available GPU Memory | Full Compatibility Report → |
| NVIDIA H100 80GB SXM | 80 GB (76 GiB) | 3350 GB/s | 204.19 GiB | 212.81 GiB | ~4.3 tok/s | Exceeds Available GPU Memory | Full Compatibility Report → |
| NVIDIA H20 96GB HBM3 | 96 GB (91.2 GiB) | 4000 GB/s | 204.19 GiB | 212.81 GiB | ~4.9 tok/s | Exceeds Available GPU Memory | Full Compatibility Report → |
| NVIDIA H20 141GB HBM3e | 141 GB (133.9 GiB) | 4800 GB/s | 204.19 GiB | 212.81 GiB | ~8.1 tok/s | Exceeds Available GPU Memory | Full Compatibility Report → |
| NVIDIA H200 141GB SXM | 141 GB (133.9 GiB) | 4800 GB/s | 204.19 GiB | 212.81 GiB | ~8.1 tok/s | Exceeds Available GPU Memory | Full Compatibility Report → |
| NVIDIA B200 180GB | 180 GB (171 GiB) | 8000 GB/s | 204.19 GiB | 212.81 GiB | ~19.7 tok/s | Exceeds Available GPU Memory | Full Compatibility Report → |
| NVIDIA B300 288GB Blackwell Ultra | 288 GB (273.6 GiB) | 8000 GB/s | 204.19 GiB | 212.81 GiB | ~141.6 tok/s | Fits with Safe Headroom | Full Compatibility Report → |
| Huawei Ascend 910B 64GB (昇腾) | 64 GB (60.8 GiB) | 1200 GB/s | 204.19 GiB | 212.81 GiB | ~3.7 tok/s | Exceeds Available GPU Memory | Full Compatibility Report → |
| Huawei Ascend 910C 128GB (昇腾) | 128 GB (121.6 GiB) | 2400 GB/s | 204.19 GiB | 212.81 GiB | ~6.6 tok/s | Exceeds Available GPU Memory | Full Compatibility Report → |
| Huawei Atlas 300I Duo 96GB (310P) | 96 GB (91.2 GiB) | 408 GB/s | 204.19 GiB | 212.81 GiB | ~4.1 tok/s | Exceeds Available GPU Memory | Full Compatibility Report → |
| Cambricon MLU590 80GB (寒武纪思元) | 80 GB (76 GiB) | 2048 GB/s | 204.19 GiB | 212.81 GiB | ~4.2 tok/s | Exceeds Available GPU Memory | Full Compatibility Report → |
| Hygon DCU K100-AI 64GB (海光深算) | 64 GB (60.8 GiB) | 1638 GB/s | 204.19 GiB | 212.81 GiB | ~3.8 tok/s | Exceeds Available GPU Memory | Full Compatibility Report → |
| MetaX MetaX C500 64GB (沐曦曦云) | 64 GB (60.8 GiB) | 1800 GB/s | 204.19 GiB | 212.81 GiB | ~3.8 tok/s | Exceeds Available GPU Memory | Full Compatibility Report → |
| Kunlunxin Kunlunxin P800 64GB (昆仑芯) | 64 GB (60.8 GiB) | 1600 GB/s | 204.19 GiB | 212.81 GiB | ~3.8 tok/s | Exceeds Available GPU Memory | Full Compatibility Report → |
| Moore Threads MTT S4000 48GB (摩尔线程) | 48 GB (45.6 GiB) | 768 GB/s | 204.19 GiB | 212.81 GiB | ~3.3 tok/s | Exceeds Available GPU Memory | Full Compatibility Report → |
| AMD Radeon RX 7900 XTX 24GB | 24 GB (22.8 GiB) | 960 GB/s | 204.19 GiB | 212.81 GiB | ~3 tok/s | Exceeds Available GPU Memory | Full Compatibility Report → |
| AMD Radeon AI PRO R9700 32GB | 32 GB (30.4 GiB) | 640 GB/s | 204.19 GiB | 212.81 GiB | ~3.1 tok/s | Exceeds Available GPU Memory | Full Compatibility Report → |
| AMD Ryzen AI Max+ 395 128GB | 128 GB (99.8 GiB) | 256 GB/s | 204.19 GiB | 212.81 GiB | ~4.4 tok/s | Exceeds Available GPU Memory | Full Compatibility Report → |
| AMD Instinct MI300X 192GB | 192 GB (182.4 GiB) | 5300 GB/s | 204.19 GiB | 212.81 GiB | ~31.8 tok/s | Exceeds Available GPU Memory | Full Compatibility Report → |
| AMD Instinct MI325X 256GB | 256 GB (243.2 GiB) | 6000 GB/s | 204.19 GiB | 212.81 GiB | ~115.1 tok/s | Fits with Safe Headroom | Full Compatibility Report → |
| AMD Instinct MI355X 288GB | 288 GB (273.6 GiB) | 8000 GB/s | 204.19 GiB | 212.81 GiB | ~141.6 tok/s | Fits with Safe Headroom | Full Compatibility Report → |
| Intel Arc Pro B60 24GB | 24 GB (22.8 GiB) | 456 GB/s | 204.19 GiB | 212.81 GiB | ~2.9 tok/s | Exceeds Available GPU Memory | Full Compatibility Report → |
| Intel Gaudi 3 128GB HBM2e | 128 GB (121.6 GiB) | 3700 GB/s | 204.19 GiB | 212.81 GiB | ~6.7 tok/s | Exceeds Available GPU Memory | Full Compatibility Report → |
| Apple Mac mini M4 Pro 64GB | 64 GB (49.9 GiB) | 273 GB/s | 204.19 GiB | 212.81 GiB | ~3.3 tok/s | Exceeds Available GPU Memory | Full Compatibility Report → |
| Apple MacBook Pro M4 Max 128GB | 128 GB (99.8 GiB) | 546 GB/s | 204.19 GiB | 212.81 GiB | ~5.5 tok/s | Exceeds Available GPU Memory | Full Compatibility Report → |
| Apple Mac Studio M4 Ultra 256GB | 256 GB (199.7 GiB) | 819 GB/s | 204.19 GiB | 212.81 GiB | ~22 tok/s | Tight Fit (<8% Headroom) | Full Compatibility Report → |
| Apple Mac Studio M3 Ultra 512GB | 512 GB (399.4 GiB) | 819 GB/s | 204.19 GiB | 212.81 GiB | ~22 tok/s | Fits in Unified Memory (Bandwidth-Bound) | Full Compatibility Report → |