What are the VRAM requirements for GLM-5.2 753B?

GLM-5.2 753B requires approximately 486.5 GB of VRAM at Q4_K_M quantization, 855.1 GB at Q8, and 1698.4 GB at FP16. These numbers assume 8k context window; VRAM scales linearly with context length due to the KV cache.

How many parameters does GLM-5.2 753B have?

GLM-5.2 753B has 753 billion total parameters, but only 40 billion are active per token thanks to its Mixture of Experts (MoE) architecture. This makes inference significantly faster than the total parameter count suggests.

How capable is GLM-5.2 753B?

GLM-5.2 753B achieves an MMLU-Pro score of 80.6, placing it among the most capable open-weight models available — competitive with frontier systems on general knowledge and reasoning.

Can GLM-5.2 753B run on a 16 GB GPU?

No. At Q4_K_M, GLM-5.2 753B needs 486.5 GB of VRAM — more than 16 GB. You will need a multi-GPU server.

Can GLM-5.2 753B run on a 24 GB GPU?

No. Even at Q4_K_M, GLM-5.2 753B needs 486.5 GB. Consider a multi-GPU server with 80 GB+ total VRAM.

What is the smallest quantization for GLM-5.2 753B that fits in 24 GB of VRAM?

GLM-5.2 753B cannot fit in 24 GB of VRAM at any standard quantization level. The minimum needed is 289.2 GB at Q2_K.

What GPU do I need to run GLM-5.2 753B locally?

You need a multi-GPU server. At Q4_K_M, GLM-5.2 753B needs 486.5 GB VRAM, more than any single consumer GPU. Consider 2–4× H100 or A100 GPUs.

GLM-5.2 753B

Name: GLM-5.2 753B
Author: Z.ai

GLM-5.2 753B needs roughly 486.5 GB VRAM at Q4_K_M quantization (1698.5 GB at FP16). 3 GPUs we track can run it fully in VRAM at 8k context.

3 GPUs run this natively · 0 with CPU offload

Z.ai753B params40B active (MoE)977k contextMITCommercial use ok

GLM-5.2 753B is a Mixture of Experts (MoE) model with 753B total parameters but only 40B active per token developed by Z.ai. Released June 13, 2026 as Z.ai's successor to GLM-5.1. Same 753B-parameter MoE class (40B active) but introduces IndexShare, a new sparse-attention mechanism that extends context from GLM-5.1's ~200K to a full 1,000,000 tokens while keeping long-context inference cost manageable. Ships in BF16 and FP8 under an MIT license.

To run GLM-5.2 753B locally: Weight footprint is the same tier as GLM-5.1 — Q2_K lands around 250-300GB, still datacenter-scale hardware (multi-GPU 80GB-class servers or comparable). The 1M-token context window is the bigger practical constraint: even with IndexShare's sparse attention, KV cache at long sequence lengths adds substantial VRAM overhead on top of the base weight requirement, so realistic long-context use needs headroom well beyond the minimum weight-only estimate. As a MoE model, inference speed depends on active parameters (40B) rather than total size.

Leads open-weight models on SWE-bench Pro (62.1%) and Terminal-Bench 2.1 (81.0%), trailing only Claude Opus 4.8 on agentic coding benchmarks. Strong GPQA-Diamond performance extends GLM-5.1's reasoning gains, and the 1M-token window makes it one of the few open-weight models usable for whole-repository or long-document agentic workflows without external retrieval.

VRAM at each quantization

Assumes 8k context. KV cache grows linearly with context length.

Quant	Weights	KV cache	Total
FP32	3012.0 GB	10.47 GB	3385.2 GB
BF16	1506.0 GB	10.47 GB	1698.5 GB
FP16	1506.0 GB	10.47 GB	1698.5 GB
Q8_0	753.0 GB	10.47 GB	855.1 GB
Q6_K	617.5 GB	10.47 GB	703.3 GB
Q5_K_M	484.9 GB	10.47 GB	554.9 GB
Q4_K_M	423.9 GB	10.47 GB	486.5 GB
Q3_K_M	323.8 GB	10.47 GB	374.4 GB
Q2_Krec	247.7 GB	10.47 GB	289.2 GB
NVFP4cuda	376.5 GB	10.47 GB	433.4 GB

KV cache shown at 8k context (FP16). NVFP4 requires a CUDA GPU. Enable TurboQuant in the calculator to see reduced KV cache estimates.

Benchmarks

MMLU-Pro

80.6

GPQA Diamond

91.2

GPUs that run GLM-5.2 753B natively (3)

Apple M4 Ultra (384GB)Q2_K · 91.3 t/s
Apple M3 Ultra (512GB)Q3_K_M · 52.4 t/s
Apple M2 Ultra (384GB)Q2_K · 66.9 t/s

Notes

1M-token context via new IndexShare sparse attention. #1 open-weight on SWE-bench Pro (62.1%) and Terminal-Bench 2.1 (81.0%), just behind Claude Opus 4.8.

Hugging Face ↗Released 2026-06-13

Compare GLM-5.2 753B with other models

How to run GLM-5.2 753B locally

816244880160320

Q2_K needs 289.2 GB — needs multiple datacenter-class GPUs (80 GB+ each).

Ollama

ollama run glm-5.2

llama.cpp

./llama-cli -m glm-5.2.Q2_K.gguf -c 32768 -ngl 99

LM Studio: GLM-5.2 is a 753B MoE model -- even at Q2 it needs roughly 250-300 GB of VRAM/unified memory. Look for pre-split GGUF quants in LM Studio if running across multiple GPUs, and note that using the full 1M-token context multiplies KV-cache VRAM needs well beyond the base weight footprint.

Why this quantization? GLM-5.2 has 753B total parameters in a Mixture-of-Experts architecture with only 40B active per token, so -- as with all large MoE models -- the full parameter count must be resident in VRAM regardless of how few experts fire per forward pass. Q2 is the practical floor that makes local inference feasible at all on multi-GPU hardware; anything higher pushes total memory well past what all but datacenter-class rigs can offer.

Who is GLM-5.2 753B for?

Teams and researchers with multi-GPU servers (4+ high-VRAM GPUs, 80 GB-class) who need frontier-level open-weight coding/agentic performance with very long context, and want a self-hosted alternative to closed long-context coding models.

Best for

Long-horizon agentic coding and multi-step software engineering (SWE-bench Pro: 62.1%)
Terminal/tool-use agent workflows (Terminal-Bench 2.1: 81.0%)
Whole-repository or long-document reasoning using the full 1M-token context
Graduate-level science and reasoning tasks (strong GPQA-Diamond performance)
Self-hosted replacement for closed long-context coding assistants

Not ideal for

Any setup with less than 4x 80 GB-class GPUs -- even at Q2 the model needs roughly 250-300 GB for weights alone
Long-context use on constrained hardware -- the 1M-token window adds KV-cache overhead on top of the base weight requirement
Latency-sensitive interactive use -- MoE routing and long-context attention both add overhead versus smaller dense models
Budget-conscious or single-GPU deployments -- consider a smaller dense coding model instead

GPUs mentioned

Apple M4 Ultra (384GB)384 GB

AppleWorkstation1092 GB/s

Apple M3 Ultra (512GB)512 GB

AppleWorkstation819 GB/s

Apple M2 Ultra (384GB)384 GB

AppleWorkstation800 GB/s

Frequently asked questions

What are the VRAM requirements for GLM-5.2 753B?: GLM-5.2 753B requires approximately 486.5 GB of VRAM at Q4_K_M quantization, 855.1 GB at Q8, and 1698.4 GB at FP16. These numbers assume 8k context window; VRAM scales linearly with context length due to the KV cache.
How many parameters does GLM-5.2 753B have?: GLM-5.2 753B has 753 billion total parameters, but only 40 billion are active per token thanks to its Mixture of Experts (MoE) architecture. This makes inference significantly faster than the total parameter count suggests.
How capable is GLM-5.2 753B?: GLM-5.2 753B achieves an MMLU-Pro score of 80.6, placing it among the most capable open-weight models available — competitive with frontier systems on general knowledge and reasoning.
Can GLM-5.2 753B run on a 16 GB GPU?: No. At Q4_K_M, GLM-5.2 753B needs 486.5 GB of VRAM — more than 16 GB. You will need a multi-GPU server.
Can GLM-5.2 753B run on a 24 GB GPU?: No. Even at Q4_K_M, GLM-5.2 753B needs 486.5 GB. Consider a multi-GPU server with 80 GB+ total VRAM.
What is the smallest quantization for GLM-5.2 753B that fits in 24 GB of VRAM?: GLM-5.2 753B cannot fit in 24 GB of VRAM at any standard quantization level. The minimum needed is 289.2 GB at Q2_K.
What GPU do I need to run GLM-5.2 753B locally?: You need a multi-GPU server. At Q4_K_M, GLM-5.2 753B needs 486.5 GB VRAM, more than any single consumer GPU. Consider 2–4× H100 or A100 GPUs.