Z.ai · mixture of experts · GLM-5.3 License

What runs GLM-5.3

Z.ai's flagship MoE from August 2026; the card does not state its parameter count. Local use needs 512 GB-class memory even when quantized.

Parameters: 753.3B
4-bit file: 435.2 GB; UD-Q4_K_XL
Needs at 8k context: 437.4 GB; context memory estimated
Longest context: 1,048,576 tokens

Context memory: 89,856 bytes per token, an estimate from the config: MLA: (512 + 64) x 78 layers x 2 bytes. Multi-head latent attention (kv_lora_rank 512, RoPE dim 64) with DeepSeek-style sparse attention: the cache stores a compressed latent per layer, so the GQA formula does not apply. Card gives no per-token figure. How context memory works.

Measured on this hardware

Machines that hold the 4-bit file with 8,192 tokens of context and have published runs on the same chip or cards, fastest first. Median of single-user runs with plain decoding.

No published run on a machine that fits it.

Fits, no speed figure

Too large for

Files

Quant	Format	Size	Repo
UD-IQ1_S	GGUF	201.8 GB	unsloth/GLM-5.3-GGUF
UD-IQ1_M	GGUF	212.8 GB	unsloth/GLM-5.3-GGUF
UD-IQ2_M	GGUF	222.2 GB	unsloth/GLM-5.3-GGUF
UD-Q2_K_XL	GGUF	236.4 GB	unsloth/GLM-5.3-GGUF
UD-IQ3_XXS	GGUF	262.3 GB	unsloth/GLM-5.3-GGUF
UD-Q3_K_XL	GGUF	319.4 GB	unsloth/GLM-5.3-GGUF
4bit	MLX	389.6 GB	mlx-community/GLM-5.3-4bit
UD-Q4_K_XL	GGUF	435.2 GB	unsloth/GLM-5.3-GGUF
Q8_0	GGUF	746.3 GB	unsloth/GLM-5.3-GGUF

Sources

huggingface.co: model card: parameters, context length, license and attention layout; release month
huggingface.co: license and parameter count from the published weights (the card gives none)
huggingface.co: layer count, KV heads and head size
huggingface.co: GGUF file sizes (shards summed)
huggingface.co: MLX 4bit safetensors sizes