GLM-4.5-Air VRAM Requirements
GLM-4.5-Air needs about 62.1 GB of VRAM at Q4_K_M with an 8K context: 59.8 GB for weights, 1.4 GB for the KV cache and 0.9 GB of runtime overhead. At Q8_0 it needs 107 GB and at full FP16 precision 200 GB. The smallest GPU that runs it fully is the A100 80GB.
62.1 GB VRAM
GLM-4.5-Air at Q4_K_M with 8K context and the KV cache in FP16.
- Weights 59.8 GB
- KV cache 1.4 GB
- Runtime overhead 0.9 GB
Mixture of experts: all 106B parameters need memory, but only 12B are read per token, so it runs at the speed of a much smaller model.
Which GPUs can run it
- RTX 3060 12GBToo small
- RTX 4060 Ti 16GBToo small
- RTX 3090 24GB36% on GPU~7.6 tok/s
- Radeon RX 7900 XTX 24GB36% on GPU~7.6 tok/s
- RTX 4090 24GB36% on GPU~7.6 tok/s
- RTX 5090 32GB48% on GPU~9.4 tok/s
- RTX 6000 Ada 48GB73% on GPU~15 tok/s
- A100 80GBFits~94 tok/s
- H100 80GB SXMFits~121 tok/s
- Ryzen AI Max+ 395 128GBFits~18 tok/s
- Apple M4 Max 128GBFits~37 tok/s
- Apple M5 Max 128GBFits~40 tok/s
- RTX PRO 6000 Blackwell 96GBFits~87 tok/s
- NVIDIA DGX Spark 128GBFits~19 tok/s
- Apple M3 Ultra 512GBFits~51 tok/s
Multi-GPU: about 3× 24 GB cards (RTX 3090 / 4090) with the model split across them.
GLM-4.5-Air VRAM by quantization
Total memory with an FP16 KV cache. "Smallest tier" is the smallest standard VRAM size with a 6% safety margin at 8K.
| Quantization | Bits/weight | Weights | Total at 8K | Total at 32K | Smallest tier |
|---|---|---|---|---|---|
| Q4_K_M | 4.85 | 59.8 GB | 62.1 GB | 66.5 GB | 80 GB |
| Q5_K_M | 5.69 | 70.2 GB | 72.5 GB | 76.8 GB | 80 GB |
| Q6_K | 6.59 | 81.3 GB | 83.6 GB | 87.9 GB | 96 GB |
| Q8_0 | 8.50 | 105 GB | 107 GB | 111 GB | 128 GB |
| FP16 / BF16 | 16.00 | 197 GB | 200 GB | 204 GB | Multi-GPU / server |
| Q3_K_M | 3.91 | 48.2 GB | 50.5 GB | 54.9 GB | 64 GB |
| Q2_K | 3.00 | 37.0 GB | 39.3 GB | 43.6 GB | 48 GB |
VRAM by context length
At Q4_K_M. The KV cache grows linearly with context: 0.18 GB per 1,000 tokens in FP16.
| Context | KV cache (FP16) | KV cache (Q8_0) | Total, FP16 cache | Total, Q8_0 cache |
|---|---|---|---|---|
| 4K tokens | 0.7 GB | 0.4 GB | 61.4 GB | 61.1 GB |
| 8K tokens | 1.4 GB | 0.8 GB | 62.1 GB | 61.5 GB |
| 32K tokens | 5.8 GB | 3.1 GB | 66.5 GB | 63.8 GB |
| 128K tokens | 23.0 GB | 12.2 GB | 83.7 GB | 72.9 GB |
GPUs that can run GLM-4.5-Air
At Q4_K_M with 8K context. "Best quant" is the highest-quality format that fits fully; "max context" is at Q4_K_M.
| GPU | Usable memory | Status | Best quant | Max context | Speed |
|---|---|---|---|---|---|
| A100 80GB | 80.0 GB | Fits | Q5_K_M | 80K | ~94 tok/s |
| H100 80GB SXM | 80.0 GB | Fits | Q5_K_M | 80K | ~121 tok/s |
| Ryzen AI Max+ 395 128GB | 96.0 GB | Fits | Q6_K | 128K | ~18 tok/s |
| Apple M3 Max 128GB | 96.0 GB | Fits | Q6_K | 128K | ~28 tok/s |
| Apple M4 Max 128GB | 96.0 GB | Fits | Q6_K | 128K | ~37 tok/s |
| Apple M5 Max 128GB | 96.0 GB | Fits | Q6_K | 128K | ~40 tok/s |
| RTX PRO 6000 Blackwell 96GB | 96.0 GB | Fits | Q6_K | 128K | ~87 tok/s |
| NVIDIA DGX Spark 128GB | 112 GB | Fits | Q6_K | 128K | ~19 tok/s |
| NVIDIA H200 141GB | 141 GB | Fits | Q8_0 | 128K | ~139 tok/s |
| Apple M2 Ultra 192GB | 144 GB | Fits | Q8_0 | 128K | ~50 tok/s |
| AMD Instinct MI300X 192GB | 192 GB | Fits | Q8_0 | 128K | ~136 tok/s |
| Apple M3 Ultra 512GB | 384 GB | Fits | FP16 / BF16 | 128K | ~51 tok/s |
| RTX 6000 Ada 48GB | 48.0 GB | 73% on GPU | Q2_K | – | ~15 tok/s |
| RTX A6000 48GB | 48.0 GB | 73% on GPU | Q2_K | – | ~15 tok/s |
| L40S 48GB | 48.0 GB | 73% on GPU | Q2_K | – | ~15 tok/s |
| A100 40GB | 40.0 GB | 61% on GPU | – | – | ~12 tok/s |
| RTX 5090 32GB | 32.0 GB | 48% on GPU | – | – | ~9.4 tok/s |
| Radeon AI PRO R9700 32GB | 32.0 GB | 48% on GPU | – | – | ~8.7 tok/s |
Multi-GPU: about 3× 24 GB cards (RTX 3090 or 4090) run it at Q4_K_M, using tensor or layer splitting in llama.cpp, vLLM or ExLlamaV2.
About GLM-4.5-Air
A 106B mixture-of-experts with 12B active parameters that became a favourite for 96–128 GB machines. It uses standard grouped-query attention across 46 layers, so a long context adds memory steadily: 184 KB per token in FP16. Strong at coding and tool use for its active size.
| Spec | Value |
|---|---|
| Parameters | 106B total, 12B active (MoE) |
| Layers | 46 |
| Hidden size | 4096 |
| Attention heads / KV heads | 96 / 8 |
| Head dimension | 128 |
| Vocabulary | 151,552 |
| Max context (used here) | 128K tokens |
| KV cache per token (FP16) | 184 KB |
| Released | 2025-07 |
| License | MIT |
GLM-4.5-Air VRAM: common questions
Can GLM-4.5-Air run on 64GB of RAM and a 24GB GPU?
Yes with partial offload. Keep attention and shared layers on the GPU and the experts in system RAM; with only 12B active parameters it still reaches usable speed.
How much VRAM does GLM-4.5-Air need?
About 62.1 GB at Q4_K_M with an 8K context, 107 GB at Q8_0 and 200 GB at FP16. Longer contexts add 0.18 GB per 1,000 tokens with an FP16 cache.
What is the cheapest GPU that can run GLM-4.5-Air?
The smallest listed device that fits it fully at Q4_K_M with 8K context is the A100 80GB (80.0 GB usable). With 24 GB cards you need about 3 of them.
Does GLM-4.5-Air need VRAM for all 106B parameters?
Yes. All experts must be loaded because any token can use any expert. Only 12B parameters are read per token, so it generates text about as fast as a 12B dense model.