Can't hold GLM locally? Rent it. GLM-4.5-Air needs one 80GB card (from ~$0.81/hr on-demand); GLM-5 needs two H200s (from ~$5.27/hr). The card maths and dated prices for each version, checked Nov 12, 2026.
Some links on this page are referral links: we may earn a commission, at no cost to you.
Share
If your machine can't hold GLM, renting is the cheapest way to run it — and the maths is just 'how much memory does this quant need, and how many cards is that.' On Nov 12, 2026 an 80GB A100 rented from about $0.81/hour on-demand (Lium; Vast.ai from $0.84) and an H200 from about $2.63/hour (Vast.ai), so even GLM-5 on a multi-GPU rental is a few dollars an hour. Here's exactly what to rent for each GLM, with the card maths shown. Part of our cloud-GPU pricing pillar.
What to rent for each GLM
The rule is simple: take the quant's memory footprint, divide by the card's memory, round up. Prices below are the cheapest offers in our tracker on Nov 12, 2026 (17:20 UTC; Vast's A100s re-read at 18:04 UTC and split by 40GB and 80GB, since Vast lists both under one name). Multi-GPU rates are per-card × count; a two- or four-card box from one host can cost more than the cheapest single cards. Runpod figures are Secure Cloud, the tier a new account sees (the Sep 24 version of this table used Community Cloud prices).
Cheapest cloud setup per GLM · checked Nov 12, 2026
GLM-4.5-Air (~60GB)
Cards needed
1× A100 80GB
Vast.ai interruptible
from $0.18/hr
Cheapest on-demand
$0.81/hr (Lium); Vast $0.84
Runpod on-demand
$1.79/hr
GLM-4.6 (~135GB)
Cards needed
1× H200 (tight), or 2× H100 80GB
Vast.ai interruptible
from $1.32/hr (1×H200)
Cheapest on-demand
$2.63/hr (1×H200, Vast); 2×H100 $4.40
Runpod on-demand
$3.79/hr (1×H200 NVL)
GLM-5 (~241GB)
Cards needed
2× H200 (~282GB)
Vast.ai interruptible
~$2.63/hr
Cheapest on-demand
~$5.27/hr (Vast)
Runpod on-demand
~$7.58/hr (2×H200 NVL)
Cards needed
Vast.ai interruptible
Cheapest on-demand
Runpod on-demand
GLM-4.5-Air (~60GB)
1× A100 80GB
from $0.18/hr
$0.81/hr (Lium); Vast $0.84
$1.79/hr
GLM-4.6 (~135GB)
1× H200 (tight), or 2× H100 80GB
from $1.32/hr (1×H200)
$2.63/hr (1×H200, Vast); 2×H100 $4.40
$3.79/hr (1×H200 NVL)
GLM-5 (~241GB)
2× H200 (~282GB)
~$2.63/hr
~$5.27/hr (Vast)
~$7.58/hr (2×H200 NVL)
Spot vs on-demand, and the multi-GPU catch
Vast.ai's spot market is cheapest but interruptible — fine for GLM-4.5-Air inference you can checkpoint. For GLM-4.6 and GLM-5 you need two to four cards at once, and multi-GPU spot capacity from a single host is rarer, so an on-demand multi-GPU instance (or a high-reliability host) is usually more practical — budget accordingly. Either way you only pay while the box is on, so a few hours with GLM-5 costs a handful of dollars, not a hardware purchase.
Or skip the GPU: GLM by the token
If you want GLM's answers rather than the weights, you don't need to rent cards at all. Z.ai, GLM's maker, sells its own API, and Novita AI serves the same models. Per 1M tokens, read on Oct 29, 2026 from Z.ai's pricing page and Novita's public model list:
GLM-4.5-Air
Z.ai official, input / output
not on Z.ai's current price list
Novita AI, input / output
$0.13 / $0.85
GLM-4.6
Z.ai official, input / output
$0.60 / $2.20
Novita AI, input / output
$0.55 / $2.20
GLM-5
Z.ai official, input / output
$1.00 / $3.20
Novita AI, input / output
$1.00 / $3.20
GLM-5.3 (newest)
Z.ai official, input / output
$1.40 / $4.40
Novita AI, input / output
$1.40 / $4.40
Z.ai official, input / output
Novita AI, input / output
GLM-4.5-Air
not on Z.ai's current price list
$0.13 / $0.85
GLM-4.6
$0.60 / $2.20
$0.55 / $2.20
GLM-5
$1.00 / $3.20
$1.00 / $3.20
GLM-5.3 (newest)
$1.40 / $4.40
$1.40 / $4.40
The prices are nearly identical, and Z.ai's cached input is cheaper ($0.26 per 1M on GLM-5.3). Pick Novita if you also want Kimi, Qwen, DeepSeek or Llama under the same key, or need GLM-4.5-Air by the token. Rent a GPU when your prompts must stay on hardware you control.
Frequently asked
What's the cheapest way to run GLM in the cloud?
Match the quant to the fewest cards. GLM-4.5-Air (~60GB at 4-bit) needs one 80GB card — from about $0.81/hour on-demand (Lium A100 80GB; Vast from $0.84) on Nov 12, 2026, and less on interruptible offers. GLM-4.6 (~135GB) needs two 80GB cards or one H200; GLM-5 (~241GB) needs two H200s or four 80GB cards. Spot is cheapest but interruptible, so checkpoint; you pay only while the instance runs.
How many GPUs do I need to run GLM-5 in the cloud?
GLM-5's 2-bit quant is about 241GB, so you need roughly 282GB of GPU memory — two H200s (141GB each) or four 80GB cards (A100/H100). Multi-GPU spot capacity is limited, so an on-demand multi-GPU instance is often the practical choice. At Vast.ai's cheapest on-demand rate on Nov 12, 2026 ($2.63 per H200) that's roughly $5.27/hour for a 2×H200 setup.
Is it cheaper to rent or buy hardware for GLM?
For occasional use, rent — an evening with GLM costs a few dollars versus a large-memory machine costing thousands. Buying a 128GB unified-memory box makes sense only if you run GLM-4.5-Air constantly; for the bigger versions, renting multi-GPU by the hour avoids owning hardware you'd rarely fill. See our rent-vs-buy math.
