How much VRAM does GLM-5.3 755B (MoE) need at Q4_K_M? Zhipu's 1M context MoE

By TechCompare · Updated · How we check

GLM-5.3 755B at Q4_K_M is a ~617 GB resident deployment for 1M context, with 461 GB of it in weights. For most buyers the Z.ai API is the better starting point than self-hosting.

Calculator

Estimated VRAM required

617 GB

755B params at Q4_K_M, 1,048,576 token context, batch 1, inference.

Weights
461 GB
KV cache
99.9 GB
Overhead
56.1 GB

Estimate accuracy: Weights within ~3% of most GGUF files from Q4_K_M up, ~7% for Q3_K_M and Q2_K. KV cache within ~5% for standard GQA models, ~10% for MLA or hybrid state estimates. Real VRAM may vary with framework (vLLM vs llama.cpp vs Transformers), Flash Attention, and driver overhead.

Hardware that fits

Unified memory machines show a model budget, not their installed memory. We hold back 8 GB for the system, or use the vendor's GPU limit when that's lower. It's a planning estimate, not a hard limit.

No single card or machine in our catalog has enough memory. The multi-GPU setups below do.

Multi-GPU setups

4× NVIDIA B200
Datacenter · 155 GB per card
720 GB
87% used
8× RTX PRO 6000 96GB
Pro · 78.0 GB per card
768 GB
82% used

Pooled cards show up once a model outgrows the biggest single card of that kind. Each card past the first keeps about 1 GB for its own buffers, and that's counted here. Your inference engine has to support the split, and speed also depends on how the cards connect.

GLM-5.3 755B at Q4_K_M with native 1M context needs about 617 GB of VRAM with all experts resident. Z.ai's August 2026 release keeps the 753B-scale pool of GLM-5.2 and the same 40B active per-token budget, with the MLA attention stack carried over from 5.2. Those 40B active params don't shrink what has to stay loaded, because different tokens pick different experts.

How this is calculated

The 755B pool at Q4_K_M is 461 GB of weights. The MLA key-value cache across 78 layers at a 512-dim latent is 100 GB at the full 1M context. Activation and runtime overhead adds about 57 GB. That makes GLM-5.3 nearly identical to GLM-5.2 on footprint at the same context, since the cache layout is unchanged.

Verdict

Use Z.ai's official API unless data residency or burst-proof capacity is the reason to self-host. Resident mode needs 8 cards of 80 GB or 5 H200 cards of 141 GB with NVLink, and capping context to 256K cuts the cache to ~25 GB.

More GLM scenarios

Kimi K3.1 Pro at Q4_K_M
A 2.78T expert pool with 24 latent attention layers and fixed state in the other 69.
View details
Mistral Large 4 at Q4_K_M
A 1T expert pool with 63 latent cache owners and 44 layers that also store an index key.
View details
Muse Spark Open at Q4_K_M
Only 16 of its 62 layers grow with context, and the other 46 hold a 2,048 token window.
View details

Frequently asked questions

How does GLM-5.3 differ from GLM-5.2 on VRAM?
GLM-5.3 is 755B total versus GLM-5.2's 753B, with the same 40B active per token and the same 78-layer MLA attention layout. The resident totals are nearly identical: ~617 GB vs ~616 GB at Q4_K_M and 1M context. Plan the same hardware for either one.
Can I run GLM-5.3 on two 80 GB GPUs?
Not with every expert loaded. Two 80 GB cards hold 160 GB, and the 461 GB of Q4 weights alone are well past that. Moving expert weights to system RAM cuts VRAM, but the host then has to hold them and every token gets slower, so measure real memory use before counting on two cards.