GLM 4.5V
107.7B total, 14.3B active · 64k context · MIT · released 2025-08-11
visionreasoningagenticchat
In short
GLM 4.5V is an open model from Z.ai with 107.7B parameters (14.3B active per token). At Q4_K_M, GLM 4.5V needs about 61 GB of memory at 8k context and 66 GB at 32k, counting the weights, the KV cache and runtime buffers. The cheapest hardware in Canada that runs GLM 4.5V fully in memory at 32k context is the NVIDIA RTX PRO 6000 Blackwell, $3,133 at Amazon, at about 46 tok/s (estimated).
Updated . Speeds are estimates; how they are worked out.
What to run it on in Canada
Fully in memory at 32k context, priced at retailers today.
Cheapest
NVIDIA RTX PRO 6000 Blackwell46tok/sest.Q4_K_M
$3,133
Best value
NVIDIA RTX PRO 6000 Blackwell46tok/sest.Q4_K_M
$3,133
Fastest
NVIDIA RTX PRO 6000 Blackwell46tok/sest.Q4_K_M
$3,133
All hardware that runs it
Everything we price, best value first. Two cards split the model by layers.
| Hardware | Quant | Fits | Decode | Price | tok/s per 1,000 |
|---|---|---|---|---|---|
| NVIDIA RTX PRO 6000 Blackwell | Q4_K_M | fits | 46 tok/s | $3,133 | 14.6 |
| GMKtec EVO-X2 Ryzen AI Max+ 395 128GB | Q4_K_M | fits | 6 tok/s | $5,099 | 1.1 |
| ASUS Ascent GX10 128GB | Q8_0 | fits | 5 tok/s | $8,299 | 0.6 |
| Mac Studio M4 Max 128GB | Q4_K_M | fits | 15 tok/s | – | – |
| Framework Desktop Ryzen AI Max+ 395 128GB | Q4_K_M | fits | 6 tok/s | – | – |
| Mac Studio M3 Ultra 512GB | Q8_0 | fits | 15 tok/s | – | – |
| NVIDIA DGX Spark 128GB | Q8_0 | fits | 5 tok/s | – | – |
| Mac Studio M3 Ultra 256GB | Q8_0 | fits | 15 tok/s | – | – |
| Mac Studio M3 Ultra 96GB | Q4_K_M | fits | 22 tok/s | – | – |
31 other cards and machines we price cannot hold it at this context.
Memory needed
Weights, KV cache and runtime buffers at each context length (KV cache at fp16).
| Quant | File | @8k | @32k | @64k |
|---|---|---|---|---|
| Q4_K_Mgguf | 59.3 GB | 61.4 GB | 65.9 GB | 71.8 GB |
| Q8_0gguf | 105.8 GB | 108.0 GB | 112.4 GB | 118.4 GB |
Run it
Copy a command for the runtime you use.
- Q4_K_M
- Ollama
ollama run hf.co/ggml-org/GLM-4.5V-GGUF:Q4_K_M - LM Studio
lms get ggml-org/GLM-4.5V-GGUF@q4_k_m - llama.cpp
llama-server -hf ggml-org/GLM-4.5V-GGUF:Q4_K_M
- Ollama
- Q8_0
- Ollama
ollama run hf.co/ggml-org/GLM-4.5V-GGUF:Q8_0 - LM Studio
lms get ggml-org/GLM-4.5V-GGUF@q8_0 - llama.cpp
llama-server -hf ggml-org/GLM-4.5V-GGUF:Q8_0
- Ollama
Notes
MoE model built on GLM-4.5-Air (106B total, 12B active); needs transformers>=4.57.1, sglang>=0.5.3 or vllm>=0.10.2, and the example serve commands use 4-way tensor parallelism. Thinking mode is on by default, and the card says pure-text Q&A is weaker because this release focused on multimodal use.
Questions
- How much VRAM does GLM 4.5V need?
- At Q4_K_M, GLM 4.5V needs about 61 GB of memory at 8k context and 66 GB at 32k, counting the weights, the KV cache and runtime buffers.
- What is the best GPU for running GLM 4.5V locally?
- The cheapest hardware in Canada that runs GLM 4.5V fully in memory at 32k context is the NVIDIA RTX PRO 6000 Blackwell, $3,133 at Amazon, at about 46 tok/s (estimated).
- Can I run GLM 4.5V on a 16 GB GPU?
- No: it does not fit in 16 GB at 32k context, even with system RAM.
- Can I run GLM 4.5V on a 24 GB GPU?
- No: it does not fit in 24 GB at 32k context, even with system RAM.
- How do I run GLM 4.5V locally?
- With Ollama: ollama run hf.co/ggml-org/GLM-4.5V-GGUF:Q4_K_M (the Q4_K_M build). LM Studio and llama.cpp commands are on this page.
- What licence is GLM 4.5V released under?
- MIT, which allows commercial use.