Model × hardware check
Can GLM-4.5-Air run on NVIDIA DGX Spark · 128 GB?
Yes, this is a comfortable estimate at 8,192 tokens. a 106B-total, 12B-active MoE that trades a large stored expert set for a smaller active path DGX Spark provides 128 GB of unified system memory and 273 GB/s of bandwidth in a compact CUDA-capable machine.
Short answer
Comfortable fit
At least 20% memory headroom remains after weights, KV cache, and runtime overhead.
- Available shared CPU/GPU memory
- 128 GB
- Estimated requirement
- 86.6 GB
- Headroom
- +41.4 GB
- Fit ratio
- 148%
- Recommended quantization
- Q5_K_M
- Context ceiling
- 131,072 tokens
- Speed estimate
- Estimate · 15.9–21.7 tok/s
Estimate · decode · llama.cpp/MLX · 8,192-token context This is a bandwidth model, not a measured benchmark.
What the numbers say
The selected GLM-4.5-Air package uses 83.5 GB of weight files. The 8,192 tokens KV cache estimate is 1.5 GB, and the formula adds 1.5 GB of runtime overhead. That produces a 86.6 GB requirement against 128 GB of shared CPU/GPU memory.
This compact AI system uses unified system memory for CPU and GPU work. It is not a conventional add-in GPU, and runtime support must be checked for the exact software stack.
The context ceiling shown above is the point where the same formula reaches the machine's capacity. It is not a promise that the backend can allocate that full context after the operating system and other applications take their share.
Quantization trade-offs at 8,192 tokens
| Quantization | Weight files | 8,192 tokens requirement | Fit on this machine | Context ceiling | Speed estimate | Ollama |
|---|---|---|---|---|---|---|
| Q3_K_M | 57.20 GB | 60.2 GB | Comfortable fit · 212% | 131,072 tokens | Estimate · 23.2–31.6 tok/s | No verified tag |
| UD-Q4_K_XL | 67.72 GB | 70.8 GB | Comfortable fit · 181% | 131,072 tokens | Estimate · 19.6–26.7 tok/s | No verified tag |
| Q5_K_M | 83.52 GB | 86.6 GB | Comfortable fit · 148% | 131,072 tokens | Estimate · 15.9–21.7 tok/s | No verified tag |
| Q8_0 | 117.46 GB | 120.5 GB | Tight fit · 106% | 47,994 tokens | Estimate · 11.3–15.4 tok/s | No verified tag |
For MoE models, a tight CPU-offload result means the weights fit in system memory even when they do not fit in accelerator memory. That does not increase the accelerator context ceiling.
Run commands
lms get "GLM-4.5-Air"@q5_k_m
lms load <model-key-from-lms-ls> --gpu max --context-length 8192llama-cli -hf unsloth/GLM-4.5-Air-GGUF:Q5_K_M -c 8192 -ngl 99 -p "Your prompt"mlx_lm.generate --model mlx-community/GLM-4.5-Air-4bit --prompt "Your prompt" --max-tokens 8192Commands are examples, not a guarantee that the selected quantization tag is available in every runtime. Confirm the package name and license on the linked model page.
Public speed checks
No verified measurement yet. The estimate above is deliberately labeled as an estimate rather than a benchmark.
If it does not fit
This pairing already clears the comfort threshold
The current estimate leaves +41.4 GB. A longer context or another GPU application can reduce that margin, so verify free memory before relying on it.
Low-cost setup upgrades
Useful upgrades around NVIDIA DGX Spark · 128 GB
These are optional setup improvements, not a claim that an accessory can increase GPU or unified-memory capacity.
External NVMe SSD for model storage
Keeps large GGUF and MLX files off the boot drive. Fast storage reduces load time, but it does not increase the memory available for inference.
Search current listings on AmazonDesktop RAM upgrade kit
Useful for MoE CPU offload and keeping the operating system out of VRAM. Match the motherboard's memory type, slot layout, capacity limit, and validated part numbers.
Search current listings on AmazonGPU support bracket
A lower-cost mechanical upgrade for heavy three-slot cards. Check clearance, mounting height, and whether the bracket blocks fans or case panels.
Search current listings on AmazonAs an Amazon Associate I earn from qualifying purchases. Links are sponsored search links and do not change the price you pay. Affiliate disclosure.
Where this estimate can break
- 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.
- Other GPU applications, display buffers, and the operating system reduce usable memory.
- Longer prompts, concurrent requests, and a different backend can raise peak memory.
Compare this pairing
FAQ
Questions about this pairing.
How much memory does GLM-4.5-Air need on NVIDIA DGX Spark · 128 GB?
At 8,192 tokens with Q5_K_M, the estimate is 86.6 GB: 83.5 GB of weights, 1.5 GB of KV cache, and 1.5 GB of runtime overhead. The available shared CPU/GPU memory is 128 GB.
What context length can NVIDIA DGX Spark · 128 GB handle with this model?
The calculated ceiling is 131,072 tokens for this quantization, capped by the model's native 131,072-token context. Actual usable context can be lower after the OS, other applications, and backend workspace.
How does quantization change the memory requirement on NVIDIA DGX Spark · 128 GB?
At 8,192 tokens, the published variants calculate as follows: Q3_K_M: 60.2 GB, comfortable fit; UD-Q4_K_XL: 70.8 GB, comfortable fit; Q5_K_M: 86.6 GB, comfortable fit; Q8_0: 120.5 GB, tight fit. A lower-bit file can move a result from too large to tight, but quality and kernel support may differ.
What can I run if GLM-4.5-Air does not fit on NVIDIA DGX Spark · 128 GB?
This selection already clears the comfort threshold. If you later increase the context, monitor the ceiling of 131,072 tokens; smaller alternatives include GLM-5.3-Flash, GLM-4.7-Flash, Flint 40B.
Is the speed estimate for GLM-4.5-Air on NVIDIA DGX Spark · 128 GB measured?
The site's Estimate · 15.9–21.7 tok/s range is an estimate from 12B active parameters, 273 GB/s accelerator memory bandwidth, and the selected quant at 8,192 tokens. It is not a benchmark. No public measurement is attached to this exact runtime and context.