What are the VRAM requirements for GLM-5.3 753B?

GLM-5.3 753B requires approximately 515.1 GB of VRAM at Q4_K_M quantization, 898.0 GB at Q8, and 1688.2 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-5.3 753B have?

GLM-5.3 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.

Can GLM-5.3 753B run on a 16 GB GPU?

No. At Q4_K_M, GLM-5.3 753B needs 515.1 GB of VRAM, more than 16 GB. You will need a multi-GPU server.

Can GLM-5.3 753B run on a 24 GB GPU?

No. Even at Q4_K_M, GLM-5.3 753B needs 515.1 GB. Consider a multi-GPU server with 516 GB+ of combined VRAM.

What is the smallest quantization for GLM-5.3 753B that fits in 24 GB of VRAM?

GLM-5.3 753B cannot fit in 24 GB of VRAM at any standard quantization level. The minimum needed is 322.8 GB at Q2_K.

What GPU do I need to run GLM-5.3 753B locally?

You need a multi-GPU server. At Q4_K_M, GLM-5.3 753B needs 515.1 GB VRAM, more than any single consumer GPU. That's roughly 7x 80 GB datacenter GPUs (H100, A100, or similar) pooled together.

Is GLM-5.3 a new architecture, or the same model as GLM-5.2?

Same architecture and same base model. GLM-5.3's config.json is identical to GLM-5.2's (78 layers, 256 routed experts, the same Multi-head Latent Attention shape), and Z.ai's own model card says every capability gain comes from post-training alone. The weights themselves changed; the VRAM math did not.

Why were GLM-5.3's open weights delayed?

Z.ai launched the model through its API on August 14, 2026, but held the open-weight release back roughly two weeks for a safety review, publishing on Hugging Face around August 25. The company says post-training produced an unplanned jump in the model's ability to chain together multi-stage exploitation plans, an offensive-security capability it hadn't specifically trained for, which is what prompted the review.

Is GLM-5.3 the same as GLM-5.3-Flash?

No, despite the shared version number. GLM-5.3 is this text-only post-training refresh of GLM-5.2's existing base model with no architecture change. GLM-5.3-Flash is a separate, newly trained model: a smaller 320B/18B-active checkpoint with a new hybrid linear/sparse-attention design and native image and video input, neither of which GLM-5.3 has.

GLM-5.3 753B

Name: GLM-5.3 753B
Author: Z.ai

GLM-5.3 753B needs roughly 515.1 GB VRAM at Q4_K_M quantization (1688.2 GB at FP16). 2 GPUs we track can run it fully in VRAM at 8k context.

2 GPUs run this natively · 0 with CPU offload

Z.ai753B params40B active (MoE)1024k contextGLM-5.3Commercial use ok

Q2_K total: 322.8 GB; at 8k context
Smallest GPU: 512 GB; Apple M5 Ultra (512GB), at Q2_K
KV cache, full context: 167.5 GB; 78 of 78 layers cache
Inputs: Text; GLM-5.3, released August 2026
Active experts: 8 routed + 1 shared; of 256 total

GLM-5.3 753B is a Mixture of Experts (MoE) model with 753B total parameters but only 40B active per token developed by Z.ai. API launched 14 August 2026; open weights held back about two weeks and published 25 August 2026 under a new, non-MIT GLM-5.3 License. Reuses GLM-5.2's exact base model and Multi-head Latent Attention shape (78 layers, 753B total, 40B active): Z.ai's own model card says every capability gain here comes from post-training alone, with no architecture change.

To run GLM-5.3 753B locally: Identical weight footprint and KV-cache math to GLM-5.2: Q2_K lands around 250-300GB, still datacenter-scale hardware (multi-GPU 80GB-class servers or comparable), and the 1,048,576-token context window's MLA-compressed KV cache adds real but manageable overhead on top of that. The post-training refresh changes capability, not what hardware can run it. As a MoE model, inference speed depends on active parameters (40B) rather than total size.

Z.ai reports a 50% jump on its in-house Z.ai Code Bench and, on its own comparison table against Kimi K3, DeepSeek-V4 Pro-0813, and closed frontier models, the best score among open-weight peers on Terminal-Bench 3.0 (28.3 vs GLM-5.2's 4.6) and Agents' Last Exam, plus Terminal-Bench 2.1 88.2 (vs 81.0). GPQA Diamond 91.7%, independently measured by Artificial Analysis since Z.ai doesn't publish it. The same post-training run produced an unplanned jump in offensive-security ability, more than doubling GLM-5.2 on ExploitBench (54.4 vs 24.4) and reaching 84.5 on CyberGym, which is what triggered the pre-release safety review.

VRAM at each quantization

Calculated at 8k context. Since KV cache scales linearly with context, longer sessions need more VRAM than shown here.

Quant	Weights	KV cache	Total
FP32	3012.0 GB	1.31 GB	3374.9 GB
BF16	1506.0 GB	1.31 GB	1688.2 GB
FP16	1506.0 GB	1.31 GB	1688.2 GB
Q8_0	800.4 GB	1.31 GB	898.0 GB
Q6_K	618.2 GB	1.31 GB	693.9 GB
Q5_K_M	536.1 GB	1.31 GB	601.9 GB
Q4_K_M	458.6 GB	1.31 GB	515.1 GB
Q3_K_M	362.2 GB	1.31 GB	407.1 GB
Q2_Krec	286.9 GB	1.31 GB	322.8 GB
NVFP4cuda	376.5 GB	1.31 GB	423.1 GB

KV cache is calculated at 8k context (FP16). Note that NVFP4 only runs on CUDA GPUs. Turn on TurboQuant in the calculator above for lower KV cache estimates.

Benchmarks

GPQA Diamond

91.7

Terminal-Bench 2.1

88.2

What five months of post-training bought, no architecture change

GLM-5.3 shares GLM-5.2's exact base model and weights shape; every benchmark gain here comes from further post-training alone, per Z.ai's own model card. These are the two largest gains on Z.ai's own benchmark table among suites with no field in this site's schema, and one of them is what pushed Z.ai to delay the open-weight release for a safety review.

ExploitBench

GLM-5.3

54.4

GLM-5.2

24.4

Terminal-Bench 3.0

GLM-5.3

28.3

GLM-5.2

4.6

Z.ai, GLM-5.3 model card benchmark table (huggingface.co/zai-org/GLM-5.3, 25 August 2026). Both scores are Z.ai's own vendor-reported figures; independent reproduction was not yet available at time of writing.

ExploitBench more than doubles, 24.4 to 54.4, and Terminal-Bench 3.0, a newer, harder suite than the Terminal-Bench 2.1 this site tracks as a structured field, rises roughly sixfold, 4.6 to 28.3. Z.ai says the ExploitBench-class gain traces back to an emergent, unplanned jump in the model's ability to chain multi-stage exploitation plans, discovered during post-training, which is why it held the open weights back for roughly two weeks of safety review before publishing them.

GPUs that run GLM-5.3 753B natively (2)

Apple M5 Ultra (512GB)Q3_K_M · 14.7 t/s
Apple M3 Ultra (512GB)Q3_K_M · 10 t/s

Notes

Same base model and weights shape as GLM-5.2; Z.ai's own model card says every gain here comes from post-training alone, with no architecture change. Z.ai reports a 50% improvement on its in-house Z.ai Code Bench and, on its own comparison table against Kimi K3, DeepSeek-V4 Pro-0813, Qwen3.8-Max, and closed frontier models, the best score among open-weight peers on Terminal-Bench 3.0 (28.3 vs GLM-5.2's 4.6) and Agents' Last Exam, plus Terminal-Bench 2.1 88.2 (vs 81.0). The same post-training run produced an unplanned jump in offensive-security ability: Z.ai says the model started chaining multi-stage exploitation plans it wasn't trained to produce, so it held the open-weight release back about two weeks for a safety review before publishing, during which it says the model surfaced 2,436 vulnerabilities across 269 open-source projects. Structured scores: CyberGym 84.5 (vs 77.2) and more than double GLM-5.2 on ExploitBench (54.4 vs 24.4). The API launched August 14, 2026; weights followed under a new, non-MIT GLM-5.3 License once that review finished.

Hugging Face ↗Released 2026-08-25

Compare GLM-5.3 753B with other models

How to run GLM-5.3 753B locally

816244880160320

Q2_K needs 322.8 GB: needs multiple datacenter-class GPUs (80 GB+ each).

Ollama

ollama run glm-5.3

llama.cpp

./llama-cli -m glm-5.3.Q2_K.gguf -c 32768 -ngl 99

LM Studio: GLM-5.3 has the same 753B MoE footprint as GLM-5.2, since it reuses the same base model and weights shape: 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 1,048,576-token context multiplies KV-cache VRAM needs well beyond the base weight footprint.

Why this quantization? GLM-5.3 has the same 753B total parameters in a Mixture-of-Experts architecture with only 40B active per token as GLM-5.2, so 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-5.3 753B for?

Teams and researchers with multi-GPU servers (4+ high-VRAM GPUs, 80 GB-class) who already run GLM-5.2 and want its post-training gains in long-horizon coding and offensive-security research without changing hardware, or anyone choosing between the two who values GLM-5.3's stronger Terminal-Bench 3.0 and CyberGym scores over GLM-5.2's MIT license.

Best for

Long-horizon agentic coding and multi-step software engineering (Terminal-Bench 2.1: 88.2%)
Offensive-security research and vulnerability discovery (CyberGym: 84.5%)
Whole-repository or long-document reasoning using the full 1,048,576-token context
Graduate-level science and reasoning tasks (GPQA-Diamond 91.7%, independently measured)
Self-hosted replacement for closed long-context coding assistants, on hardware that already runs GLM-5.2

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
Deployments that specifically require an MIT license: GLM-5.3's is a custom license, permissive but not MIT
Long-context use on constrained hardware, the 1,048,576-token window adds KV-cache overhead on top of the base weight requirement
Budget-conscious or single-GPU deployments, consider a smaller dense coding model instead

Should you run this instead of GLM-5.2?

If you already have the multi-GPU, 80GB-class hardware GLM-5.2 needs (roughly 250-300GB at this site's recommended Q2 quant, unchanged here since the weights share the same shape), GLM-5.3 is a straightforward upgrade for coding and agentic work: real, Z.ai-reported gains on Terminal-Bench 3.0 and CyberGym, at no extra VRAM cost over GLM-5.2. The trade is licensing, not hardware: GLM-5.2 ships under a plain MIT license, while GLM-5.3 uses a new, custom (though still broadly permissive) GLM-5.3 License. Nobody below the 80GB-class multi-GPU tier gets more from this release than from GLM-5.2. Given the model's demonstrated exploitation-chaining ability, treat it like any other offensive-security-capable tool: appropriate for authorized research and testing, not casual or unsupervised deployment.

Explore

Browse all guides →Browse all models →Browse all GPUs →

Frequently asked questions

What are the VRAM requirements for GLM-5.3 753B?: GLM-5.3 753B requires approximately 515.1 GB of VRAM at Q4_K_M quantization, 898.0 GB at Q8, and 1688.2 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-5.3 753B have?: GLM-5.3 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.
Can GLM-5.3 753B run on a 16 GB GPU?: No. At Q4_K_M, GLM-5.3 753B needs 515.1 GB of VRAM, more than 16 GB. You will need a multi-GPU server.
Can GLM-5.3 753B run on a 24 GB GPU?: No. Even at Q4_K_M, GLM-5.3 753B needs 515.1 GB. Consider a multi-GPU server with 516 GB+ of combined VRAM.
What is the smallest quantization for GLM-5.3 753B that fits in 24 GB of VRAM?: GLM-5.3 753B cannot fit in 24 GB of VRAM at any standard quantization level. The minimum needed is 322.8 GB at Q2_K.
What GPU do I need to run GLM-5.3 753B locally?: You need a multi-GPU server. At Q4_K_M, GLM-5.3 753B needs 515.1 GB VRAM, more than any single consumer GPU. That's roughly 7x 80 GB datacenter GPUs (H100, A100, or similar) pooled together.
Is GLM-5.3 a new architecture, or the same model as GLM-5.2?: Same architecture and same base model. GLM-5.3's config.json is identical to GLM-5.2's (78 layers, 256 routed experts, the same Multi-head Latent Attention shape), and Z.ai's own model card says every capability gain comes from post-training alone. The weights themselves changed; the VRAM math did not.
Why were GLM-5.3's open weights delayed?: Z.ai launched the model through its API on August 14, 2026, but held the open-weight release back roughly two weeks for a safety review, publishing on Hugging Face around August 25. The company says post-training produced an unplanned jump in the model's ability to chain together multi-stage exploitation plans, an offensive-security capability it hadn't specifically trained for, which is what prompted the review.
Is GLM-5.3 the same as GLM-5.3-Flash?: No, despite the shared version number. GLM-5.3 is this text-only post-training refresh of GLM-5.2's existing base model with no architecture change. GLM-5.3-Flash is a separate, newly trained model: a smaller 320B/18B-active checkpoint with a new hybrid linear/sparse-attention design and native image and video input, neither of which GLM-5.3 has.