Skip to content
ReviewRamble

The weekly brief

AI, reviewed, in your inbox.

  • New reviews of tools, models, MCP servers and repos, with the verdict up front
  • Re-checked prices and ratings when something changes
  • New analysis pieces as they are published

Models, agents, MCP servers and the tools actually worth your time.

One email a week. Unsubscribe with the link in any brief. How we use your email

GLM 4.5 Air GPU requirements: VRAM and cheapest GPU

GLM 4.5 Air has 110.5B parameters, published in BF16. It is a mixture-of-experts model: 8 of its 128 experts work on each token, about 17B parameters, so it generates at roughly the speed of a model that size. Every expert still has to sit in GPU memory, which is why the memory figures below are for the full model. It is released under mit, which allows commercial use.

Parameters
110.5B
Smallest footprint (INT4)
about 66 GB
Cheapest GPU per hour (FP16)
$0.59
Type
Text
  • Parameters 110.5B
  • Active per token 17B
  • Type Text
  • Task text-generation
  • Max context 128K
  • Licence mit
  • Downloads 1.1M
  • Likes 650

Calculating for GLM 4.5 Air

Precision
Workload
More settings
Serving engine
Requests at once

GLM 4.5 Air needs about 246 GB of GPU memory

Inference (vLLM), FP16 / BF16, 8K tokens context, one request at a time.

Where the memory goes

  • Weights 221 GB
  • KV cache 1.5 GB
  • Runtime overhead 23 GB
  • Total 246 GB

GPU counts keep 10% of each card free for spikes. Plan extra if you expect long prompts or many users at once.

GPUs that fit, cheapest firstEvery GPU type we track that can hold this setup on one node, ranked by total hourly cost
RankConfigurationPer GPU / hrGPUs neededTotal / hrTotal / month (24/7)Provider
16 × A40 48GBCheapestAmpere · GDDR6$0.596sold as 8 GPUs; you need 6$4.72$3,446RunPod (opens in a new tab)8x A40
22 × MI300X 192GBCDNA 3 · HBM3$2.392$4.78$3,489RunPod (opens in a new tab)2x MI300X
34 × A100 80GBAmpere · HBM2e$1.304$5.18$3,781Vast.ai (opens in a new tab)4x A100 SXM4 (median of 6 offers)
46 × RTX A6000 48GBAmpere · GDDR6$0.696sold as 8 GPUs; you need 6$5.52$4,030RunPod (opens in a new tab)8x RTX A6000
53 × RTX PRO 6000 96GBBlackwell · GDDR7$1.733sold as 4 GPUs; you need 3$6.94$5,065Vast.ai (opens in a new tab)4x RTX PRO 6000 S (median of 9 offers)
62 × H200 141GBHopper · HBM3e$3.792$7.58$5,533RunPod (opens in a new tab)2x H200 NVL
71 × B300 288GBBlackwell Ultra · HBM3e$8.001$8.00$5,840Lium (opens in a new tab)1x B300 SXM6
86 × L40S 48GBAda Lovelace · GDDR6$1.096sold as 8 GPUs; you need 6$8.72$6,366RunPod (opens in a new tab)8x L40S
96 × RTX 6000 Ada 48GBAda Lovelace · GDDR6$1.096sold as 8 GPUs; you need 6$8.72$6,366RunPod (opens in a new tab)8x RTX 6000 Ada
104 × H100 80GBHopper · HBM3$2.894$11.56$8,439RunPod (opens in a new tab)4x H100 PCIe
113 × H100 NVL 94GBHopper · HBM3$3.193sold as 4 GPUs; you need 3$12.76$9,315RunPod (opens in a new tab)4x H100 NVL
122 × B200 180GBBlackwell · HBM3e$7.892$15.78$11,519Lambda (opens in a new tab)2x B200 SXM6
137 × A100 40GBAmpere · HBM2$1.997sold as 8 GPUs; you need 7$15.92$11,622Lambda (opens in a new tab)8x HGX A100
142 × GB200 186GBBlackwell · HBM3e$10.502sold as 4 GPUs; you need 2$42.00$30,660CoreWeave (opens in a new tab)GB200 NVL72
More than 8: needs several nodes: RTX 5090 32GB, RTX 4090 24GB, RTX 3090 24GB, L4 24GB, A10 24GB, T4 16GB, V100 16GB
Cost per cloudCheapest offer that fits, per provider. Approximate, prices collected Nov 16, 2026.
ProviderInstancePer GPU / hrGPUs neededTotal / hrTotal / month (24/7)
RunPod (opens in a new tab)Cheapest8x A40us$0.596 × A40 48GB sold as 8 GPUs; you need 6about $4.72about $3,446
Vast.ai (opens in a new tab)4x A100 SXM4 (median of 6 offers)worldwide marketplace$1.304 × A100 80GB about $5.18about $3,781
Lium (opens in a new tab)1x B300 SXM6worldwide marketplace$8.001 × B300 288GB about $8.00about $5,840
Nebius (opens in a new tab)1x B300 288GBeu/us$9.501 × B300 288GB about $9.50about $6,935
Vultr (opens in a new tab)vbm-48c-1024gb-4-a100-gpu (bare metal)us$2.404 × A100 80GB about $9.59about $7,000
Azure (opens in a new tab)NC96ads_A100_v4eastus$3.674 × A100 80GB about $14.69about $10,725
Lambda (opens in a new tab)2x B200 SXM6us$7.892 × B200 180GB about $15.78about $11,519
AWS (opens in a new tab)g7e.24xlargeus-east-1$4.143 × RTX PRO 6000 96GB sold as 4 GPUs; you need 3about $16.57about $12,098
Google Cloud (opens in a new tab)g4-standard-192us-central1$4.503 × RTX PRO 6000 96GB sold as 4 GPUs; you need 3about $18.00about $13,140
CoreWeave (opens in a new tab)L40Sus$2.256 × L40S 48GB sold as 8 GPUs; you need 6about $18.00about $13,140

Prices collected Nov 16, 2026. Every cost is approximate: on-demand, US dollars, mostly US regions (marketplaces such as Vast.ai and Lium are worldwide).

Memory by precision and workload

With a 8K-token context and one request at a time. LoRA and full fine-tuning start from 16-bit weights and QLoRA from 4-bit, so the other cells are blank.
PrecisionInferenceLoRA fine-tuneQLoRA fine-tuneFull fine-tune
FP16 / BF16246 GB288 GB–1,950 GB
INT8130 GB–––
INT466 GB–108 GB–

Cheapest GPU by precision

Inference with an 8K context and one request at a time.

FP16 / BF16

about 246 GB · 8K tokens

The weights as published. Nothing lost, but the largest footprint.

6 × A40 48GB

RunPod (opens in a new tab) · 8x A40

sold as 8 GPUs; you need 6
Per GPU / hr
$0.59
GPUs needed
6
Total / hr
about $4.72
per month, 24/7, approx.
$3,446

INT8

about 130 GB · 8K tokens

Weights at half the size. Most chat and RAG work shows no visible difference.

4 × A40 48GB

RunPod (opens in a new tab) · 4x A40

Per GPU / hr
$0.59
GPUs needed
4
Total / hr
about $2.36
per month, 24/7, approx.
$1,723

INT4

about 66 GB · 8K tokens

Weights at a quarter of the size. Small drops show up first in maths and code.

1 × RTX PRO 6000 96GB

RunPod (opens in a new tab) · 1x RTX PRO 6000 MaxQ

Per GPU / hr
$0.50
GPUs needed
1
Total / hr
about $0.50
per month, 24/7, approx.
$365

Inference vs fine-tuning

  • Inference (FP16)

    1×

    of the memory inference needs

    246 GB

    6 × A40 48GB · about $4.72

  • LoRA fine-tune

    1.2×

    of the memory inference needs

    288 GB

    7 × A40 48GB · about $4.72

  • QLoRA fine-tune

    0.4×

    of the memory inference needs

    108 GB

    3 × A40 48GB · about $2.36

  • Full fine-tune

    7.9×

    of the memory inference needs

    1,950 GB

    8 × B300 288GB · about $71.92

Serving it at FP16 fits on 6 × A40 48GB (RunPod). LoRA needs 7 × A40 48GB (RunPod), QLoRA 3 × A40 48GB (RunPod) and a full fine-tune 8 × B300 288GB (RunPod). Training keeps activations and optimiser state in memory as well as the weights, which is why it so often needs more GPUs than serving the same model.

Will it run on a Mac or a single RTX 4090?

RTX 4090 (24 GB)

Not on its own. Even at INT4 it needs about 66 GB, more than the card’s 24 GB. Two or more cards, or a bigger data-centre GPU, will do it.

Mac with Apple silicon

Yes, with enough memory. At INT4 it needs about 66 GB, so a Mac with 96 GB or more of unified memory (macOS lets the GPU use about three quarters of it). Expect slower generation than on a data-centre GPU.

Similar models

Open models of a similar size.

Calculate another model

GLM 4.5 Air: common questions

How much VRAM does GLM 4.5 Air need?

About 246 GB at FP16, 130 GB at INT8 and 66 GB at INT4, with an 8K context and one request at a time. Longer contexts and more parallel requests add KV cache on top.

What is the cheapest GPU to run GLM 4.5 Air?

6 × A40 48GB (about $0.59 per GPU / hour; sold as 8 GPUs; you need 6) on RunPod, at about $4.72 an hour for FP16 / BF16 inference with an 8K context. Prices collected Nov 16, 2026; they change often.

Can I fine-tune GLM 4.5 Air on a single GPU?

Not on one GPU. Even QLoRA needs about 108 GB, more than an 80 GB GPU holds. LoRA needs about 288 GB and a full fine-tune about 1,950 GB.

How are these numbers calculated?

Weights are parameters times bytes per parameter (2 at FP16, 1 at INT8, 0.5 at INT4). The KV cache comes from the model’s layers, KV heads, head size and context length. We add about 10% plus 1 GB for the runtime and fill each GPU to 90%. Prices are on-demand list prices in US regions, collected from each provider’s public pricing.

Can GLM 4.5 Air run on a consumer GPU like the RTX 4090?

Not on its own. Even at INT4 it needs about 66 GB, more than the card’s 24 GB. Two or more cards, or a bigger data-centre GPU, will do it.

Does GLM 4.5 Air fit on a single H100 80GB?

Yes, at INT4 with an 8K context.

How many GPUs do I need for GLM 4.5 Air?

At FP16 with an 8K context: 4 × 80 GB, 6 × 48 GB or 12 × 24 GB GPUs. At INT4 you need about a quarter of the memory for the weights.

How much VRAM does GLM 4.5 Air need at INT8?

About 130 GB: 116 GB for the weights, 1.5 GB for an 8K KV cache and 13 GB of runtime overhead.

How much does it cost to run GLM 4.5 Air 24/7 for a month?

About $3,446 on the cheapest setup we found (6 × A40 48GB (about $0.59 per GPU / hour; sold as 8 GPUs; you need 6), $4.72 an hour for 730 hours). The cheapest on AWS, Google Cloud or Azure is about $10,725 a month.

FP16 vs INT8 vs INT4: which should I use?

FP16 keeps full quality and needs the most memory (about 246 GB here). INT8 halves the weights with little loss for most chat work. INT4 (AWQ, GPTQ or GGUF Q4) cuts them to a quarter, about 66 GB in total, with a small drop in quality that shows most in maths and code.

Can GLM 4.5 Air run on a Mac?

Yes, with enough memory. At INT4 it needs about 66 GB, so a Mac with 96 GB or more of unified memory (macOS lets the GPU use about three quarters of it). Expect slower generation than on a data-centre GPU.

vLLM or llama.cpp?

vLLM is built for serving many requests at once on NVIDIA and AMD GPUs and reserves most of the GPU memory for KV cache. llama.cpp runs GGUF files on almost anything, including Macs and CPUs, and suits one user or a small team. Memory needs are similar for the same precision.

Prices collected Nov 16, 2026. Every cost is approximate: on-demand, US dollars, mostly US regions (marketplaces such as Vast.ai and Lium are worldwide).

How we calculate

  • Weights. Parameters times bytes per parameter: 2 at FP16, 1 at INT8, 0.5 at INT4, plus about 5% for the quantisation scales at INT8 and INT4.
  • KV cache. 2 × layers × KV heads × head size × context × requests × 2 bytes, from the model’s config.json. Where a setting is missing we estimate it and say so.
  • Overhead. About 10% plus 1 GB for the runtime, in line with what vLLM reserves.
  • Fine-tuning. LoRA keeps 16-bit weights and adds a small adapter, its optimiser and activations. QLoRA does the same over 4-bit weights. A full fine-tune needs about 16 bytes per parameter with mixed-precision Adam (weights, gradients and optimiser states), plus activations.
  • GPUs. We fill each GPU to 90% of its memory and stop at 8 per node. Costs are on-demand list prices in US regions; marketplace clouds show the median offer.