Skip to content

GLM 4.5 Air on 2× RTX 5090: VRAM, speed and fit

GLM 4.5 Air at 3-bit runs on the 2× RTX 5090 with 48K of context at about 116 tokens a second.

Text generation · Chat · Ollama / LM Studio

Runs comfortably · on 2× RTX 5090 · 3-bit (Q3_K_M) · Ollama / LM Studio · Chat

GPU memory
60.4 GB
of 64.0 GB
Context
48K
auto
Speed
116 tok/s
63 when full
First reply
0.6 s
2K prompt

Where the memory goes

Model
50.3 GB
Context
8.6 GB
Overhead
1.5 GB
OS
0.0 GB
Free
3.6 GB
Memory
60.4 GB of 64.0 GB
Context
48K tokens · auto
Speed
116 tok/s · 63 when full
First reply
0.6 s for a 2K prompt
Running cost
$17/mo at home · $117/mo on a cloud 4× RTX 2000 Ada

Validation: Validated

What GLM 4.5 Air on 2× RTX 5090 was checked against
CheckTargetResultStatus
Fits in GPU memory64.0 GB usable60.4 GB neededPass
Headroom5% free3.6 GB free (6%)Pass
Context for chat16K or more48KPass
Writing speed15 tok/s or more116 tok/sPass
First replyunder 3 s, 2K prompt0.6 sPass

Targets: a reply you can read along with, a first reply the use case tolerates, and 5% of memory left free. Calculated from published specs, not benchmarked.

The setup

Model
GLM 4.5 Air
Parameters
110B, 15B active
Precision
3-bit (Q3_K_M)
Context
48K, auto
Chats at once
1
Use
Chat
Runs in
Ollama / LM Studio
Hardware
2× RTX 5090
GPU can use
64.0 GB of 64.0 GB

Worth trying

  • 8-bit cache: Room for 88K context

Running cost

At home
$17/mo
770 W at the wall
Renting instead
$117/mo
4× RTX 2000 Ada at $0.96/h

4 h a day, power at $0.18 per kWh, cloud at RunPod on-demand prices as of Nov 21, 2026.

How to run it

llama.cpp
llama-server -hf unsloth/GLM-4.5-Air-GGUF
vLLM
vllm serve zai-org/GLM-4.5-Air
MLX (Mac)
mlx_lm.chat --model mlx-community/GLM-4.5-Air-4bit

GLM 4.5 Air on similar machines

All machines: GLM 4.5 Air VRAM requirements. Everything that runs here: what can the 2× RTX 5090 run?

Other models that run on 2× RTX 5090

Questions

Can GLM 4.5 Air run on 2× RTX 5090?

GLM 4.5 Air at 3-bit runs on the 2× RTX 5090 with 48K of context at about 116 tokens a second.

Report VR-0QPBFMF, calculated Nov 22, 2026 from published specs, not benchmarked.