Zhipu GLM-4.5-Air 106B (MoE) VRAM Requirements Guide
Compact 106B MoE agentic coding and tool-use powerhouse (12B active). Designed for Dual RTX 5090 (64GB), RTX PRO 6000 (96GB), Strix Halo, and DGX Spark.
>98.2% retention. 62.6 GB weights, leaving +30.1 GB headroom on AMD Ryzen AI Max+ 395 Strix Halo 128GB.
Near-lossless 8-bit precision requiring 119.3 GB VRAM for benchmark evaluations.
Uncompressed FP16 weights taking 219.6 GB. Demands multi-GPU or datacenter accelerators.
Architectural Memory Breakdown & Attention Scaling
GLM 4.5 Air 106B is a high-efficiency Mixture-of-Experts model containing 106B total parameters with 14.5B active parameters per token across 48 layers with 8 GQA key-value heads. Requiring ~64 GB at 4-bit precision, it is designed for dual RTX 5090 (64GB) rigs, AMD Strix Halo 128GB, and workstation-scale AI deployments.
On Dual RTX 5090 or Mac Studio workstations, pass `-fa on -ctk q8_0` in llama.cpp to preserve context buffer space. This enables 32k multi-lingual context analysis within 68 GB total system VRAM.
AMD Ryzen AI Max+ 395 Strix Halo 128GB provides 128.0 GB physical VRAM (+30.1 GB operational margin at 8k context).
Run Zhipu GLM-4.5-Air 106B (MoE) Locally
ollama run glm-4-5-air-106b:q4_K_Mvllm serve Z.ai / Zhipu/glm-4-5-air-106b --max-model-len 8192 --gpu-memory-utilization 0.95Frequently Asked Questions
Can GLM 4.5 Air 106B run on an AMD Strix Halo 128GB?
Yes! Strix Halo allocates up to 96GB of unified memory to GPU compute, holding the 64 GB model checkpoint plus full 32k context with room to spare.
What languages does GLM 4.5 Air specialize in?
It features deep bidirectional training in English and Chinese, offering top-tier cross-lingual reasoning and technical documentation analysis.
Test Custom Hardware & Context Windows
Simulate Zhipu GLM-4.5-Air 106B (MoE) across all 31 GPUs, adjust context up to 131,072 tokens, or calculate RunPod break-even costs.
Open Zhipu GLM-4.5-Air 106B (MoE) in GPU Garden Matchmaker →