Mortgage LLM thread spotlights privacy-cost tradeoffs

This is less a model recommendation than a reality check for teams trying to self-host sensitive document workflows on a small-company budget.

• –The post captures a common enterprise tension: strict data control pushes teams toward local models, but usable long-context inference still drives serious GPU spend
• –Commenters split the problem by workload, pointing toward VLMs like Mistral Small 24B or Qwen3-VL for document extraction and larger text models for RAG-heavy knowledge tasks
• –The strongest advice is architectural, not just model-level: success depends on ingestion pipelines, retrieval strategy, concurrency planning, and security controls as much as raw model quality
• –One reply argues Microsoft Copilot-style enterprise offerings can be cheaper in practice once liability, upkeep, and hardware refresh cycles are factored in
• –Cloud GPU options such as Azure or GCP A100 instances emerge as a middle ground between full on-prem deployment and fully managed SaaS