We have organized models that are usable in the field based on VRAM, throughput, and Korean performance.
Model size is directly tied to VRAM. Based on 4-bit quantization, approximately the same number of GB as the model size is required.
| Model size | Q4 VRAM | Recommended GPU | Purpose |
|---|---|---|---|
| 7~8B | Approx. 5GB | L4 24GB / RTX 4060Ti | summarization, classification, RAG |
| 14B | Approx. 9GB | L4 24GB | complex document processing |
| 32B | Approx. 20GB | L4 24GB (single) | High-difficulty reasoning |
English benchmark rankings do not carry over unchanged to Korean. You must compare using actual work documents.
The largest model is not always the best. Most field work is sufficient with 7–14B.
If you tell us your site conditions, we will review a configuration that can be applied.
Technical Inquiry →