GLM-5.3-Flash VRAM requirements: every GGUF size and which GPUs fit

GLM-5.3-Flash is a mixture-of-experts (MoE) model from Z.ai (Zhipu AI). The weights went up on Hugging Face on August 25, 2026, under the MIT license, and the main repo shows about 6.3 million downloads in the last month.^[1] The model card describes 320B total parameters with 18B active per token, and calls it the first natively multimodal model in the GLM-5 series (text, image and video input).^[2] Its config.json shows 45 layers, 288 routed experts with 8 picked per token plus 1 shared expert, and a context window of 1,048,576 tokens.^[3]

Because it is an MoE, only 18B parameters work on each token, so it runs much faster than a dense 320B model. But all of the weights still have to sit in memory (or be offloaded to system RAM).

Memory needed at each quantization

Quantization stores each weight in fewer bits. Smaller files need less memory, at some cost in quality, and the cost grows below about 4 bits. Below are Z.ai's FP8 checkpoint, NVIDIA's NVFP4 build, and the GGUF builds from Unsloth and ggml-org (the llama.cpp team), using the exact file sizes on Hugging Face (1 GB = 1,000,000,000 bytes, multi-file builds summed), largest first.^[4][6][7][8] The vision projector (mmproj, 0.7 to 1.2 GB) is only needed for image or video input and is not included.^[6][7]

Z.ai's main repo is stored in FP8 (the quantization_config in config.json); a separate BF16 repo holds the full-precision weights at 642.7 GB.^[3][5] NVIDIA's NVFP4 build (204.4 GB) is meant for Blackwell GPUs. ggml-org also publishes small separate files for multi-token prediction (mtp-…, 4.9 to 9.3 GB) that you only need for speculative decoding.^[7] Builds whose names start with "UD-" are Unsloth's dynamic quants.

Which build should you pick? On a 256 GB Mac or four 80 GB GPUs, a 4-bit or 5-bit build. On two 96 GB GPUs, UD-IQ4_XS. On a 128 GB Mac, UD-Q2_K_XL, and expect a clear quality loss at 2 bits.

Context length adds memory (the KV cache)

GLM-5.3-Flash mixes two kinds of attention. Its config.json lists 34 linear-attention layers, which keep a fixed-size state instead of a per-token cache, and only 11 sparse latent-attention layers that grow with the conversation.^[3] Each of those 11 layers stores a 512-value compressed latent plus a 128-value indexer key per token. At BF16 that is about 14,080 bytes per token. Z.ai has not published a per-token figure, so treat this as our estimate; your engine may store the cache differently. The fixed state of the 34 linear layers adds roughly 71 MB per conversation at BF16, whatever the length.

So even the full 1M-token window adds about 15 GB, which is small next to the weights.

Which GPU setups and Macs it fits on

Each cell compares the weights plus an 8K-token KV cache with the total memory of the setup.^{[9][10][11][12][13]}

Column headers show weights + an 8,192-token KV cache (our estimate from config.json). Multi-GPU rows assume the model is split evenly across all GPUs (tensor parallel). NVFP4 needs Blackwell GPUs (such as the RTX PRO 6000 or B200), so ignore that column on H100/H200 and Mac.

How to read it: "Fits" means weights + 8K cache use under 90% of memory. "Tight" means 90–100%: it may load, but the engine's own buffers or a longer context can push it over. "No" means it won't fit. The 90% line is our own rule of thumb, not a vendor figure. On a Mac, macOS also needs some of the unified memory, so treat those rows as optimistic.

Our recommendations

These follow from the tables above.

How it compares with other big open MoE models

4-bit file sizes come from each model's Hugging Face GGUF repo, the same data as our comparison table.^[14]

†Unsloth UD-Q4_K_XL (no Q4_K_M build exists). GLM-5.3-Flash is ggml-org's Q4_K; the others are Q4_K_M.

How to run it

The model card lists SGLang, vLLM, TokenSpeed, Transformers, KTransformers and Unsloth as supported ways to run it locally, each with its own guide.^[2] For llama.cpp, download a GGUF build from Unsloth's or ggml-org's repo.^[6][7] The card says thinking effort is set with reasoning_effort (low, high or max, default max), and that chat apps should pass clear_thinking=true.^[2]

If the model is bigger than your VRAM, llama.cpp can keep some experts in system RAM; see our MoE expert offload guide. With 18B active parameters per token this is more usable than on the bigger models above, but still much slower than running fully on the GPU. We have not measured its speed.

Cheapest GPUs to rent for GLM-5.3-Flash

Two 96 GB RTX PRO 6000s hold the UD-IQ4_XS build; four of them hold the official FP8 weights. On-demand prices from Runpod's pricing page (checked October 31, 2026; Community Cloud is cheaper but runs on shared hosts):^[12]

Prices checked on 2026-11-01. They change often, so confirm on the provider's page. Runpod lists per-GPU prices; the multi-GPU rows are that price times the GPU count. Whether a multi-GPU pod is available at any moment varies. Rent on Runpod Referral link

See all GPU rental prices for more providers and multi-GPU setups.

Frequently asked questions

Can I run GLM-5.3-Flash on one GPU?

Not on a consumer GPU. The smallest build, Unsloth's 1-bit UD-IQ1_S, is 93.1 GB, more than any 24 GB or 32 GB card.^[6] You need two or more data-center or workstation GPUs, a Mac with 128 GB or more of unified memory, or a lot of system RAM with expert offloading.

How much VRAM does GLM-5.3-Flash need at 4-bit?

ggml-org's Q4_K GGUF is 181.4 GB, and an 8K-token context adds about 0.1 GB by our estimate, so about 181.5 GB before engine overhead.^[7][3] Unsloth's UD-Q4_K_XL is 199.7 GB.^[6] A 256 GB Mac Studio holds either; two 96 GB GPUs (192 GB) are tight for Q4_K.

Will it run on a 128 GB Mac?

Only at 2 bits or lower. UD-Q2_K_XL (108.7 GB) fits by our rule, UD-IQ3_XXS (120.4 GB) is tight, and the 4-bit builds do not fit.^[6][13]

Why is the context so cheap?

Most of its layers (34 of 45) use linear attention, which keeps a fixed-size state instead of storing every token. Only the 11 sparse latent-attention layers grow with the conversation, and they store a compressed cache.^[3]

Is it open source?

The weights are released under the MIT license, which allows commercial use.^[1]

How we calculated

• Weights are the exact file sizes on Hugging Face, checked November 1, 2026, with multi-file builds added up.^{[4][5][6][7][8]} The vision projector and MTP files are not included.
• KV cache is our estimate from config.json (11 sparse-attention layers × (512 latent + 128 indexer values) at 2 bytes; the 34 linear-attention layers keep a fixed state instead).^[3] Z.ai has not published a figure.
• Device memory is from NVIDIA, Apple and Runpod spec and pricing pages. Multi-GPU rows add up memory and assume an even split.
• Rental prices are the listed on-demand Runpod per-GPU rates from our GPU price data, checked October 31, 2026, times the GPU count.^[12]
• Not included: engine working buffers, the OS and display. These vary by engine and settings, so leave headroom.

Sources

All sources checked on 2026-11-01.

1. Hugging Face, zai-org/GLM-5.3-Flash (MIT license, repo creation date, downloads in the last month; checked November 1, 2026). https://huggingface.co/zai-org/GLM-5.3-Flash
2. GLM-5.3-Flash model card (320B total / 18B active, multimodal, hybrid attention, supported engines, reasoning settings). https://huggingface.co/zai-org/GLM-5.3-Flash/blob/main/README.md
3. Hugging Face, config.json (45 layers, 34 linear-attention and 11 sparse-attention layers, latent and indexer sizes, experts, context length, FP8 quantization config). https://huggingface.co/zai-org/GLM-5.3-Flash/blob/main/config.json
4. Hugging Face, zai-org/GLM-5.3-Flash file list (official FP8 checkpoint, 62 safetensors files). https://huggingface.co/zai-org/GLM-5.3-Flash/tree/main
5. Hugging Face, zai-org/GLM-5.3-Flash-BF16 file list (120 safetensors files). https://huggingface.co/zai-org/GLM-5.3-Flash-BF16/tree/main
6. Hugging Face, unsloth/GLM-5.3-Flash-GGUF file list. https://huggingface.co/unsloth/GLM-5.3-Flash-GGUF/tree/main
7. Hugging Face, ggml-org/GLM-5.3-Flash-GGUF file list. https://huggingface.co/ggml-org/GLM-5.3-Flash-GGUF/tree/main
8. Hugging Face, nvidia/GLM-5.3-Flash-NVFP4 file list. https://huggingface.co/nvidia/GLM-5.3-Flash-NVFP4/tree/main
9. NVIDIA, H100 (80 GB). https://www.nvidia.com/en-us/data-center/h100/
10. NVIDIA, H200 (141 GB). https://www.nvidia.com/en-us/data-center/h200/
11. NVIDIA, RTX PRO 6000 Blackwell (96 GB). https://www.nvidia.com/en-us/products/workstations/professional-desktop-gpus/rtx-pro-6000/
12. Runpod, GPU pricing (per-GPU prices; checked October 31, 2026). https://www.runpod.io/pricing
13. Apple, Mac Studio tech specs (128 GB, 256 GB and 512 GB unified memory options). https://www.apple.com/mac-studio/specs/
14. VRAMNeeded comparison table (sources listed per model). https://vramneeded.com/compare/