How to Deploy GLM 5.2 Locally

• Min VRAM at 4-bit468 GB · 8K
• Parameters744B
• Active params40B
• Max context1M

Architecture

MoE

Experts offloaded

14 GB VRAM + 455 GB RAM

License

MIT

Launch date

Jun 16, 2026

Vendor

Z.ai / Zhipu AI

Hugging Face repo

zai-org/GLM-5.2

GGUF repo

unsloth/GLM-5.2-GGUF

NVFP4 repo (Blackwell)

nvidia/GLM-5.2-NVFP4

MLX repo (Apple)

mlx-community/GLM-5.2-4bit

Layers

78

Hidden size

6,144

KV heads

64

Head dim

192

Full calculation rules are on the Methodology page.

Only common GPUs and systems are listed, one set for models under 50B and another for larger ones, sorted by price from low to high.

Device	Cards	Total price	Total VRAM	Speed t/s	Total power
DGX Spark 128GB128 GB per unit	4 units	$32,960	512 GB	7.6	960 W
Instinct MI21064 GB per card	8 cards	$39,952	512 GB	38.8	2,400 W
RTX PRO 5000 72GB72 GB per card	8 cards	$85,776	576 GB	32.9	2,400 W
A100 80GB PCIe80 GB per card	8 cards	$151,992	640 GB	44.3	2,400 W
RTX PRO 600096 GB per card	8 cards	$180,000	768 GB	41.7	4,800 W

Mac Studios link up over Thunderbolt 5 and run with MLX or exo; vLLM doesn't support them. Full calculation rules are on the Methodology page.

Detailed calculation rules are on the Methodology page.

As an eBay Partner, VRAMGlass may be compensated if you buy through the eBay links on this page. The price you pay stays the same.

1. 1

   How it differs from GLM 5.3

   Zhipu says 5.3 and 5.2 share the same base model, and all the gains come from post-training: 5.3 is clearly stronger at complex coding and long tasks, with a 50% improvement on Zhipu's internal coding benchmark. Parameters, architecture and file sizes are nearly identical, so moving from 5.2 to 5.3 doesn't require new hardware.

2. 2

   Turning off thinking is its strength

   5.2 has three modes: no thinking, High and Max. To turn thinking off, pass enable_thinking: false via chat_template_kwargs; in llama.cpp you can also just add --reasoning off. 5.3's thinking can't be turned off, only adjusted between low, high and max, so services that need fast, short answers still have a use for 5.2.

3. 3

   License

   5.2 is plain MIT, and Zhipu explicitly states there are no regional restrictions. 5.3 switched to Zhipu's own license, which adds one clause: model API providers with very large revenue must pass Zhipu's safety review before commercial use. For individuals and ordinary companies deploying it for their own use, the two licenses make no practical difference.

4. 4

   llama.cpp needs a build from late July or later

   The sparse attention in the GLM 5 series relies on an “indexer” to pick out what to look at, and that indexer only landed in mainline llama.cpp on July 24, 2026 (PR #25407). Earlier versions ignore it, and the original PR explicitly said quality suffers as a result. Unsloth GGUFs you've already downloaded don't need re-downloading; just update llama.cpp.

llama-server -hf unsloth/GLM-5.2-GGUF:Q4_K_M -c 8192

llama-server -hf unsloth/GLM-5.2-GGUF:Q4_K_M -c 8192 -ngl 99 --cpu-moe

vllm serve zai-org/GLM-5.2-FP8 \
  --kv-cache-dtype fp8 \
  --tensor-parallel-size 8 \
  --speculative-config.method mtp \
  --speculative-config.num_speculative_tokens 5 \
  --tool-call-parser glm47 \
  --reasoning-parser glm47 \
  --enable-auto-tool-choice \
  --served-model-name glm-5.2-fp8

mlx_lm.server --model mlx-community/GLM-5.2-4bit

File names follow that week's model-library snapshot; the definitions are in the methodology. Methodology

Model	Total params	Active params	4-bit size
GLM 4.7 FlashMoE	30B	3B	18.3 GB
GLM 4.5 AirMoE	106B	12B	73 GB
GLM 5.3 FlashMoE	320B	18B	199.7 GB
GLM 5.3MoE	744B	40B	467.3 GB

About GLM