Skip to content
Llama Builds

Z.ai · Released July 2025

GLM-4.5-Air

What it takes to run GLM-4.5-Air locally: memory by quantization, the GPUs that handle it, and how fast it runs.

Updated

Parameters
106B (12B active)
Architecture
Mixture of experts
Context
128K tokens
License
MIT license
Memory at Q4_K_M
~68GB

What do you need to run GLM-4.5-Air?

GLM-4.5-Air needs about 68GB of VRAM at Q4_K_M with 8K context (64GB for the weights plus KV cache and overhead). The smallest GPU that can hold it is a 72GB card like the RTX PRO 5000 72GB, with little memory to spare. For a good balance of price and speed we recommend the RTX PRO 6000 (~80 tokens/sec). Because it is a mixture-of-experts model with 12B active parameters, it generates text much faster than a dense model of the same size, and it tolerates partial CPU offload better.

See the full hardware guide for GLM-4.5-Air

What to buy

GPUs for GLM-4.5-Air

Unified memory

Ryzen AI Max+ 395

128GB unified · ~$3,300

Quant
Q5_K_M
Speed (est.)
~13 tok/s
Runs well
Check price at Amazon

Multi-GPU

3× RTX 3090

72GB total · ~$3,150 used

Quant
Q4_K_M
Speed (est.)
~38 tok/s
Fits, tight
Check price at Amazon

Single card

RTX PRO 5000 72GB

72GB · ~$9,100

Quant
Q4_K_M
Speed (est.)
~73 tok/s
Fits, tight
Check price at Amazon

Speeds are estimates at the quantization shown with 8K context. We may earn a commission when you buy through these links. Learn more.

Memory

GLM-4.5-Air VRAM requirements by quantization

Lower-bit quantizations shrink the model at some cost to quality. Q4_K_M is the usual sweet spot. Totals include the KV cache for the context length shown plus runtime overhead.

Memory needed for GLM-4.5-Air at each quantization level
QuantizationQualityWeightsTotal @ 8K ctxTotal @ 32K ctx
Q2_KRough40 GB43 GB48 GB
Q3_K_MUsable52 GB55 GB60 GB
Q4_K_MRecommendedDefault64 GB68 GB73 GB
Q5_K_MVery good76 GB80 GB85 GB
Q6_KNear-lossless87 GB92 GB97 GB
Q8_0Practically lossless113 GB118 GB123 GB
BF16Full precision212 GB221 GB225 GB

KV cache note: Standard grouped-query attention: 2 x 46 layers x 8 KV heads x 128 dims x 2 bytes = 188,416 bytes (184 KiB) per token, about 1.5GB at 8K and 24.7GB at the full 128K.

Performance

How fast GLM-4.5-Air runs

Single-user generation speed on popular GPUs that fit the model entirely in memory, at the best quantization each can hold.

  1. RTX PRO 6000

    Q5_K_M · 96GB

    ~80tok/s

  2. RTX PRO 5000 72GB

    Q4_K_M · 72GB

    ~73tok/s

  3. M5 Ultra 512GB

    Q8_0 · 512GB

    ~28tok/s

  4. M5 Ultra 256GB

    Q8_0 · 256GB

    ~28tok/s

  5. M5 Max 128GB

    Q5_K_M · 128GB

    ~25tok/s

  6. M4 Max 128GB

    Q5_K_M · 128GB

    ~23tok/s

  7. M3 Ultra 512GB

    Q8_0 · 512GB

    ~20tok/s

  8. Ryzen AI Max+ 395

    Q5_K_M · 128GB

    ~13tok/s

  9. DGX Spark

    Q6_K · 128GB

    ~12tok/s

  10. Ryzen AI Max+ PRO 495

    Q8_0 · 192GB

    ~9.3tok/s

Estimated (method)

Overview

About GLM-4.5-Air

GLM-4.5-Air is Z.ai's 106B mixture-of-experts with 12B active parameters, built for agentic coding and tool use with a switchable thinking mode. Z.ai's later releases went either much larger (the GLM-5 series) or smaller (GLM-4.7-Flash), so Air is still the GLM model aimed at 96GB to 128GB machines. It needs roughly 70GB at Q4_K_M with 8K context: a Mac or Strix Halo system with 96GB or more runs it fully in memory, and a 24GB GPU with 64GB of system RAM can run it with experts offloaded. Its full-attention KV cache is large, so long contexts cost more memory than on gpt-oss-120b or Qwen3.5 122B-A10B.

Strengths

  • +Strong agentic coding and tool calling for its size
  • +Toggleable thinking mode
  • +12B active parameters keep it responsive on unified-memory machines
  • +MIT license

Weaknesses

  • −Needs roughly 70GB at Q4_K_M
  • −Large KV cache (184 KiB per token) limits long contexts
  • −Text only
  • −Now trails gpt-oss-120b and Qwen3.5 122B-A10B on most benchmarks
CodingAgents & toolsReasoningChatToggleable reasoning

Setup

How to run GLM-4.5-Air

llama.cpp
$ llama-server -hf unsloth/GLM-4.5-Air-GGUF:Q4_K_M -c 8192 -ngl 99
MLX (Mac)
$ mlx_lm.chat --model mlx-community/GLM-4.5-Air-4bit

Weights: zai-org/GLM-4.5-Air · GGUF: unsloth/GLM-4.5-Air-GGUF

Alternatives

Models like GLM-4.5-Air

OpenAI

gpt-oss-120b

117B (5.1B active) · 128K ctx

~68GBat MXFP4
ReasoningCodingAgents & tools

Alibaba Qwen

Qwen3.5 122B-A10B

122B (10B active) · 256K ctx

~77GBat Q4_K_M
ChatReasoningCoding

Z.ai

GLM-5.3-Flash

320B (18B active) · 1.0M ctx

~201GBat Q4_K_M
CodingAgents & toolsReasoning

Meta

Llama 4 Scout

109B (17B active) · 10.5M ctx

~70GBat Q4_K_M
ChatVisionLong context

Frequently asked questions

How much VRAM do you need to run GLM-4.5-Air?

About 68GB at Q4_K_M with 8K tokens of context: 64.3GB of weights, 1.5GB of KV cache, and 2.5GB of runtime overhead. Higher-precision quants and longer contexts need more; see the table above.

What is the cheapest GPU that can run GLM-4.5-Air?

The smallest card in our database that can hold it is the RTX PRO 5000 72GB (72GB), at Q4_K_M with little memory to spare, generating ~73 tokens/sec.

Is GLM-4.5-Air free for commercial use?

GLM-4.5-Air is released under the MIT license. That is a permissive license, so commercial use is allowed.

How long a context does GLM-4.5-Air support?

Up to 131,072 tokens (128K). Long contexts need a lot of extra memory for the KV cache, so most local setups run 8K to 32K.

Sources: GLM-4.5-Air model card (Hugging Face), Unsloth GLM-4.5-Air GGUF files. Capability score basis: Placed between the Qwen3 32B and gpt-oss-120b anchors from its Artificial Analysis Intelligence Index v3 standing in reasoning mode (October 2025).