GLM-5.3-Flash (320B-A18B)
First natively multimodal GLM-5 model: 320B total, 18B active, MIT license. Thinking budget via reasoning_effort (low/high/max, default max).
Recommended quants
- Unsloth UD-Q2_K_XL (108.7 GB) / UD-IQ2_XXS (101.8 GB) / UD-IQ1_S (93.1 GB)
unsloth/GLM-5.3-Flash-GGUFThe only sizes near 128 GB machines; context must stay small. - UD-Q4_K_XL (199.7 GB)
unsloth/GLM-5.3-Flash-GGUF - Q8_0 (341 GB)
unsloth/GLM-5.3-Flash-GGUF
Chat template Works
reasoning_effort low/high/max; defaults to max if omitted or unrecognized.
Fix: Pass reasoning_effort explicitly for shorter answers.
Sampling settings
| Z.ai eval default | temp 1.0, top_p 0.95 |
|---|
Runtimes
| vLLM | Unverified | |
|---|---|---|
| llama.cpp | Works | Unsloth GGUFs. |
| Ollama | Unverified | Library page glm-5.3-flash exists; may be cloud. |
| LM Studio | Unverified | |
| MLX | Unverified |
Tool calling in agent tools
| Claude Code (local endpoint) | Unverified | |
|---|---|---|
| OpenCode | Unverified | |
| Cline | Unverified | |
| Aider | Unverified |
Known issues
Sources checked
Spotted something wrong or a new fix? Corrections are reviewed before publishing. (Submission form coming soon.)