Q4_K_M · 4K context
GLM-5.3-Flash (MoE) MoE model with 320B total parameters and 18B active per token · estimated Q4_K_M memory about 191GB
Hybrid sparse+linear attention MoE model with a total of 320B parameters (18B active). Supports the official 1,048,576 (1M) token context (text_config.max_position_embeddings) and loads in large memory environments of 256GB or more.
- Estimated Q4_K_M memory
- about 191GB
- total parameters
- 320B
- active parameter
- 18B
- native context
- 1024K
Start with a model
Find hardware for this model
Also running document search?
Reserve this for embedding and reranking models sharing the same GPU or unified memory. Leave it at none if they run separately on the CPU.
Matching devices 7· 5 shown
Comfortable fit · Q4_K_M · Headroom 45.5GB
₩19,250,000
51.2 ~ 76.8 tok/s
Comfortable fit · Q4_K_M · Headroom 41.8GB
₩19,818,800
18.5 ~ 29.6 tok/s
Comfortable fit · Q4_K_M · Headroom 289.8GB
₩25,000,000
55.5 ~ 61.6 tok/s
Comfortable fit · Q4_K_M · Headroom 289.5GB
₩38,475,000
46.1 ~ 81.9 tok/s
Can run · check speed · Q4_K_M · Headroom 0.0GB
₩14,600,000
3.4 ~ 3.4 tok/s
Price is the range midpoint, with a basic host added for GPU-only prices. Speed is an estimate for the same model and 4K input.
GLM-5.3-Flash (MoE) Local Serving Run Recipe
Selecting a device changes the quantization, safety context, GPU offload, Flash Attention, KV cache, and batch start values.
If you want to choose equipment again
You can first narrow down the model configurations based on budget, noise level, operating system and whether it's acceptable used.