OPEN WEIGHTS · MITVISION + GUI AGENT106B TOTAL · 12B ACTIVEOFFICIAL GGUF

Can you run GLM-4.5V locally?

Yes, but not on a typical 64GB machine. The official Q4_K_M GGUF is 63.6GB before runtime overhead, context, and the separate vision projector. Plan on 96GB system or unified memory for a practical Q4 start; 128GB gives safer headroom.

106B / 12B
total / active parameters
63.6GB
official Q4_K_M weights
~67.4GB+
loaded Q4 + runtime, before KV/projector
96GB
practical Q4 starting tier

The hardware verdict

These are memory-fit estimates, not measured speed claims. Vision input, context length, backend, and CPU/GPU splitting change the real working set.

64GB PC / Mac
No
The Q4 weight file nearly fills all memory before the OS and runtime load.
96GB system / Mac
Practical Q4
Enough for Q4 at moderate context, with disciplined memory use.
128GB system / Mac
Recommended Q4
Safer room for macOS, multimodal projector, context, and applications.
48GB GPU
Partial offload only
Q4 cannot fully reside in VRAM; use CPU/RAM offload or multiple GPUs.
80GB GPU
Full Q4 possible
Weights fit with useful headroom, though context and vision buffers still matter.
192GB+ system
Comfortable Q8
The 113.6GB Q8 file plus overhead fits without squeezing the OS.

Official GGUF sizes

AICanRun derives these weights from the ggml-org repository file list. Auxiliary files such as MTP and the roughly 1GB multimodal projector are intentionally excluded from the weight-size total; you still need the matching projector to use images.

QuantWeight downloadBase loaded setRealistic hardware
Q4_K_M63.6 GB~67.4 GB+96GB minimum; 128GB preferred
Q8_0113.6 GB~119.9 GB+192GB+ system memory

Why 12B active does not mean 12B memory

GLM-4.5V is a mixture-of-experts model: routing uses about 12B parameters for each token, but all 106B parameters remain stored and addressable. Active parameters help explain compute and decoding behavior; the complete GGUF determines weight memory.

Open the model compatibility page →

What GLM-4.5V is useful for

Z.ai positions GLM-4.5V for image, video, and document understanding; visual reasoning and grounding; chart and diagram analysis; and GUI-agent tasks. Local execution is most attractive when documents or screenshots must stay private, but the 106B total weight count makes it a workstation-class model.

Purpose-built high-bandwidth accelerators can change this equation. Xiaomi's recent O3 + O100 phone prototype reached 303 tokens/s with a much smaller custom MiMo 3B build; it is a useful demonstration of the bandwidth advantage, not a GLM-4.5V benchmark.

Primary sources

Z.ai model card — architecture, 106B/12B parameter counts, multimodal capabilities and MIT license ↗ggml-org GLM-4.5V GGUF — official quant files, projector and llama.cpp usage notes ↗

FAQ

How much RAM does GLM-4.5V need?

The official Q4_K_M GGUF is 63.6GB. Loaded weights and runtime need at least about 67.4GB before KV cache, image projector, vision inputs, and operating-system reserve, so 96GB system memory is the practical starting tier and 128GB is more comfortable.

Can 64GB RAM run GLM-4.5V?

Not as a reliable all-local Q4 setup. The 63.6GB Q4_K_M weight file alone nearly fills 64GB; loaded-weight overhead, the operating system, runtime buffers, KV cache, and the separate multimodal projector push it over the limit.

Can an 80GB GPU run GLM-4.5V?

An 80GB GPU can hold the 63.6GB Q4_K_M weights with meaningful headroom, but context, compute buffers, and multimodal processing still matter. A 48GB GPU cannot fully offload Q4 and needs CPU/RAM offload or multiple GPUs.

Can a 96GB or 128GB Mac run GLM-4.5V?

A 96GB Apple-silicon Mac is the practical entry point for Q4 at moderate context. A 128GB Mac leaves better room for macOS, the image projector, KV cache, and other applications. Q8 is too tight at 128GB once overhead is included.

Does 12B active mean GLM-4.5V uses only 12B-model memory?

No. MoE routing activates about 12B parameters per token, reducing compute, but the full 106B parameter set remains stored and addressable. Memory follows the full GGUF weight size, not only active parameters.