12. Hardware sizing

Profile	Hardware	Models	Notes
Production, up to about 5k invoices a month	One 48 GB GPU (L40S class)	Qwen3.8-27B FP8 plus Qwen3-Embedding-0.6B on the same card	A vendor write-up (Uvik Software) reports the FP8 build fits a single 48 GB card. Treat as indicative
Production, higher volume	Two 80 GB GPUs or four GPUs in tensor-parallel	Qwen3.8-27B full precision, long context	Month-end batch peaks
Budget	Two 24 GB consumer GPUs	Qwen3.8-27B AWQ 4-bit	Community quantizations inherit Apache-2.0, but verify each repository
CPU-only, development	16–32 GB RAM	Qwen3-8B or gpt-oss-20b, embeddings on CPU, Docling on CPU	Slow but fully open source. Development and CI only

Docling's layout model is small and runs acceptably on CPU. Give docling-serve 4–8 vCPU. Model choice is a one-line change in models.yaml.