Model record
GLM-4.5-Air
Z.ai · MIT · released · verified
One-line answer
GLM-4.5-Air needs about 60.2 GB at 8K; NVIDIA RTX PRO 6000 Blackwell · 96 GB is the smallest comfortable catalog match.
The estimate combines 57.2 GB of published weights with 1.5 GB of KV cache and 1.5 GB of runtime overhead. It is a capacity estimate, not a benchmark.
| 8K memory estimate | 60.2 GB 57.2 GB weights + 1.5 GB KV + 1.5 GB overhead |
|---|---|
| Recommended quantization | Q3_K_M Smallest published fit used for the one-line answer |
| Minimum comfortable hardware | NVIDIA RTX PRO 6000 Blackwell · 96 GB 96 GB memory target |
| Context ceiling on that target | 131,072 tokens |
What this model is
Z.ai reports 106B total parameters and 12B active parameters. The 46-layer model uses 128 routed experts, eight selected per token, plus one shared expert.
The architecture is mixture of experts. Memory must hold the full weight set, while active parameters affect the speed heuristic. Experts can be offloaded to system memory, but that trades speed for capacity.
How this GLM model actually runs locally
GLM-4.5-Air stores 106B parameters and activates about 12B, so it is much smaller than GLM-5.3 but still far larger than a 30B-A3B model. The published Q3_K_M GGUF is 57.2 GB and the Q4-class build is roughly 68-73 GB. That makes a 64 GB Mac a tight Q3 experiment, while a 96 GB GPU or Mac is the first comfortable catalog target.
Z.ai's official deployment path uses vLLM, SGLang, and Transformers. Community GGUF and MLX conversions are available, but there is no official Ollama library tag in the checked sources. Apple Silicon users must use the MLX package or a current llama.cpp build; they should not treat a generic Ollama namespace as an official file.
Runtime support was verified on 2026-11-02. Exact package names and kernel support can change quickly; use the source links below for the model, quantization, and runtime projects.
Memory and quantization
Decimal GB from Unsloth's GGUF repository. The published MLX 4-bit conversion is a separate package, and there is no official Ollama library tag.
| Quantization | Published files | 8K estimate | Smallest comfortable GPU | Smallest comfortable Mac | Context at smallest GPU / Mac | Speed estimate | Run commands |
|---|---|---|---|---|---|---|---|
| Q3_K_M | 57.20 GB | 60.25 GB | NVIDIA RTX PRO 6000 Blackwell · 96 GB · 96 GB VRAM | Mac Studio M5 Max · 128 GB · 128 GB unified memory | 131,072 / 131,072 tokens | Relative class only · Moderate relative class | lms get "GLM-4.5-Air"@q3_k_m |
| UD-Q4_K_XL | 67.72 GB | 70.76 GB | NVIDIA RTX PRO 6000 Blackwell · 96 GB · 96 GB VRAM | Mac Studio M5 Max · 128 GB · 128 GB unified memory | 131,072 / 131,072 tokens | Relative class only · Moderate relative class | lms get "GLM-4.5-Air"@ud-q4_k_xl |
| Q5_K_M | 83.52 GB | 86.56 GB | No catalog match | Mac Studio M5 Max · 128 GB · 128 GB unified memory | No catalog match | Relative class only · Moderate relative class | lms get "GLM-4.5-Air"@q5_k_m |
| Q8_0 | 117.46 GB | 120.50 GB | No catalog match | Mac Studio M5 Ultra · 256 GB · 256 GB unified memory | No catalog match | Relative class only · Moderate relative class | lms get "GLM-4.5-Air"@q8_0 |
Estimate formula: published weight files + 1.5435 GB per 8K tokens × context / 8K + 1.5 GB runtime overhead. Sizes are decimal GB, not GiB. At 8K context, the smallest listed build is about 60.2 GB before OS and application reserve.
Conservative FP16 GQA estimate using all 46 layers, eight KV heads, and a 128-wide head. Linear-attention state is not included because this model uses ordinary attention geometry.
Ollama tags were checked against the linked Ollama library pages on 2026-11-02. LM Studio's lms get suffix selects a file when the search results expose that exact quantization; confirm the local key with lms ls.
KV method source: https://huggingface.co/zai-org/GLM-4.5-Air/raw/main/config.json
Speed heuristic
The model page does not have one universal speed number: memory capacity and bandwidth belong to the machine, not the checkpoint. Open a model-and-hardware page for the bandwidth-based range and its conditions.
No universal local measurement is claimed. The 12B active path is the useful speed heuristic, while the 106B stored expert set determines memory and offload behavior.
Active path: 12B parameters. This is a relative speed signal only; it is not a tokens/sec benchmark. No verified measurement is substituted for an estimate.
Public speed checks for this model
No verified measurement yet for this exact model record.
If the full weights do not fit in accelerator memory but do fit in system memory, the calculator marks the result as a tight CPU-offload fit. Expect substantially slower generation and verify the runtime's MoE offload support.
Runtimes
- llama.cpp
- MLX
- vLLM
- SGLang
No official Ollama library tag. Use llama.cpp, MLX on Apple Silicon, or the official vLLM and SGLang paths.lms get "GLM-4.5-Air"@q3_k_m
lms load <model-key-from-lms-ls> --gpu max --context-length 8192License restrictions
MIT permits commercial use, modification, and redistribution; preserve the copyright and permission notice.
This is a plain-language summary, not legal advice. Read the linked license before commercial deployment or redistribution.
Known issues
- The model is often described as about 60 GB at 4-bit, but the published Q3_K_M file is 57.2 GB and the Q4 build is roughly 68-73 GB. Read the exact file size before choosing a 64 GB machine.
- On a 64 GB Mac, Q3_K_M is only a tight fit after KV cache and runtime overhead. A 96 GB or larger machine leaves useful headroom.
- There is no official Ollama library tag in the checked sources. Do not treat an unrelated community namespace as an official package.
- Community MLX conversions work through mlx-lm, while official Z.ai runtime guidance emphasizes vLLM and SGLang.
If it does not fit, try these smaller records
These alternatives have a lower recommended memory target in the current catalog. Open each page to compare the quantization-specific fit before downloading.
gpt-oss-120b
MoE · starting at 66 GB
Qwen3.6 35B-A3B
MoE · starting at 25 GB
Gemma 4 31B
Dense · starting at 23 GB
GLM-4.7-Flash
MoE · starting at 17 GB
FAQ
Model-specific questions
What is the smallest local file for GLM-4.5-Air?
The smallest published weight package in this record is Q3_K_M at 57.20 GB. The 8K runtime estimate adds 1.54 GB of KV cache and 1.5 GB overhead, for about 60.2 GB before operating-system reserve.
How much memory do the published quantizations need at 8K?
Q3_K_M about 60.2 GB at 8K; UD-Q4_K_XL about 70.8 GB at 8K; Q5_K_M about 86.6 GB at 8K; Q8_0 about 120.5 GB at 8K. These are calculated estimates using the published weight sizes and the model's KV-cache assumption, not live memory measurements.
Which GPU and Mac configurations are comfortable for GLM-4.5-Air?
Against the current catalog, the smallest configurations that clear the 20% headroom target are: Q3_K_M: NVIDIA RTX PRO 6000 Blackwell · 96 GB; UD-Q4_K_XL: NVIDIA RTX PRO 6000 Blackwell · 96 GB; Q5_K_M: no catalog match; Q8_0: no catalog match. A larger configuration may still be preferable for longer context or faster memory bandwidth.
What is the right Ollama or LM Studio command for GLM-4.5-Air?
There is no verified local Ollama tag for GLM-4.5-Air. Use the llama.cpp command generated for the selected quantization, or the official vLLM/SGLang path, or the linked community MLX conversion. Do not substitute an Ollama cloud tag for a downloadable local model.
What should I run instead if GLM-4.5-Air does not fit?
Smaller catalog alternatives include gpt-oss-120b, Qwen3.6 35B-A3B, Gemma 4 31B, GLM-4.7-Flash. Compare their quantization tables and minimum memory before choosing; a smaller parameter count does not always mean a proportionally smaller download.
Sources
- https://huggingface.co/zai-org/GLM-4.5-Air
- https://huggingface.co/unsloth/GLM-4.5-Air-GGUF
- https://huggingface.co/mlx-community/GLM-4.5-Air-4bit
- https://huggingface.co/lmstudio-community/GLM-4.5-Air-GGUF
- https://huggingface.co/zai-org/GLM-4.5-Air/raw/main/config.json
This page is an independent compatibility guide, not official documentation. Verify the primary source before downloading or deploying a model.