Real estate AI workflows are content-volume-heavy and visually rich: listing descriptions, market analysis, property image enhancement, virtual staging. The combination of SDXL for image work + 7B LLM for text + structured-output extraction makes this a strong fit for self-hosted SMB infrastructure.
Stack: Mistral 7B FP8 (descriptions, summaries) + SDXL + brand LoRA (image enhancement, staging) + BGE-large for property search on 4090. ~£289/mo serves an estate agency or PropTech SaaS at SMB-to-mid scale. ~£0.0001 per listing description vs ~£0.05 hosted at scale.
Workloads
- Listing descriptions: from property attributes + photos → brand-voice listing copy
- Market analysis: comparable properties + price trend summarisation
- Image enhancement: SDXL-based virtual staging, sky replacement, lighting correction
- Property search: semantic search over listing database (BGE embeddings)
- Buyer / tenant Q&A: chatbot grounded in property + neighbourhood data
Stack
- Mistral 7B FP8: descriptions, summaries, Q&A
- Brand-voice LoRA: agency-specific tone (~£10 of GPU time, real quality lift)
- SDXL + IP-Adapter: virtual staging, image enhancement
- BGE-large + Qdrant: property semantic search
- Whisper (optional): voice-to-text for property notes / inspection records
- Total VRAM: ~13 GB on 4090, comfortable
Verdict
For estate agencies + PropTech SaaS, single-4090 self-hosted covers the SMB-to-mid scale. £289/mo all-in for an entire AI tier of content generation, image enhancement, and search. Replaces ~£500-2,000/mo of typical hosted PropTech AI subscriptions.
Bottom line
4090 = real estate AI workhorse. See stock photography.
