GLM 5.3 Flash

Official model card

Parameters

320B total

Active 18B active

Context

1,048,576 tokens

From the official card

License

On the official Hugging Face card

AMDai does not reprint a license string that is not already cited on this site. Read the license on the model’s own card.

Recommended

UD-Q4_K_XL

256K · llama.cpp first

Official card: 320B total / 18B active. 11 of 45 layers are sparse attention (MLA, kv_lora_rank 512, K-only cache) and grow the context store plus an 11×384 indexer (raw + 2 pooled rows). 34 linear (KDA) layers hold a fixed state. max_position_embeddings 1,048,576. 288 routed experts + 1 shared, 8 per token. Consumer cards offload; both RAM lines are estimates until measured.

Official card: zai-org/GLM-5.3-Flash · GGUF: unsloth/GLM-5.3-Flash-GGUF. Abliterated derivatives are educational only — read /abliterated, never the default on this page.

GGUF file sizes

Q6, Q5, and Q4 sizes are cited from the GGUF repo listing. Q3 is a labeled estimate. GB here is the publisher’s decimal listing, not measured VRAM. Q1 and Q2 stay off.

System RAM

Kit sizes are real DIMM steps. Measured peaks come from run cards when we have them; otherwise the line is an estimate. Wherever RAM offload appears, both shopper labels stay on the page.

Speeds

not yet measured on our R9700s. Contribute an amdai-bench/run@1 card, or open Stack Lab doctor. Do not treat calculator GB as tok/s.

not yet measured on our R9700s. Record a card with amdai-bench or open Stack Lab doctor. Other backends on the official card stay secondary.

Configuration table

Q6 / Q5 / Q4, plus Q3 as a labeled estimate. 1 GPU and 2 GPU. Decode, prefill, peak VRAM per card, VRAM total, and peak system RAM come from amdai-bench/run@1 cards when a cell is measured. A null field stays not on the card. Untested cells say no data — contribute or not yet measured on our R9700s. Q1 and Q2 stay off.

Launch commands

llama.cpp first. Strata only where this site already documents it. Flags marked on the recipe stay shape-only. Recommended context is 256K. vLLM, Ollama, and Transformers appear on the official card as secondary paths — they are not the AMDai default.

# Shape only — not a measured local desktop recipe.
./build-vk/bin/llama-server \
  -m ~/models/UD-Q4_K_XL/GLM-5.3-Flash-UD-Q4_K_XL-00001-of-00006.gguf \
  -ngl 99 \
  -c 8192 \
  --cache-type-k f16 \
  --cache-type-v f16 \
  --host 127.0.0.1 --port 8080

Fit by GPU

Calculator VRAM at the shopper quant, 1 GPU and 2 GPUs (split). The R9700 35B one-card badge is the measured card; other SKUs stay estimates. Speed is not invented here. Offload rows still show both RAM labels on this page.

Compare with…

Head-to-head pages use the same sizes, cited context, and calculator fit already on the model pages. Q3 is a labeled estimate. No invented prices. GLM 5.3 Flash against the other cited /models pages:

• Compare models
• vs Llama 3.3 70B
• vs Qwen 3.8 Flash
• vs Qwen3.6-35B

Also: Gemma 3 27B · gpt-oss 20B · Qwen3.6 27B · Qwen3.8 27B