Skip to content
NEWLive updates feed from Hugging Face, arXiv and GitHub  See what changed →

Open-source LLMs for banking

Banks run customer service, KYC extraction, credit summarisation and policy search on open models kept inside the data centre.

Where banking teams use language models

  • Customer service assistantlow risk

    Account queries, card issues, product information across chat, app and IVR

  • KYC & onboarding document processingmedium risk

    Extract fields from ID cards, passports, utility bills; flag mismatches

  • Credit memo & loan file summarisationmedium risk

    Summarise applications and financials for analysts; no automated decisions

  • AML / fraud alert triage & SAR draftinghigh risk

    Rank alerts, draft suspicious-activity narratives for investigators

  • Regulatory & policy Q&Alow risk

    Search over central-bank circulars, internal policies and product manuals

  • Relationship-manager copilotmedium risk

    Meeting prep, call notes, next-best-action suggestions

  • Collections & dispute correspondencemedium risk

    Draft compliant reminders and dispute responses

  • Complaint classification & routinglow risk

    Categorise and prioritise complaints against ombudsman SLAs

  • Trade-finance document checkingmedium risk

    Letters of credit, bills of lading: consistency and discrepancy checks

  • Core-banking / COBOL code assistancelow risk

    Explain, document and modernise legacy code

  • Regulatory report draftinghigh risk

    First drafts of periodic returns from structured data

  • Security-operations assistantmedium risk

    Summarise incidents, draft runbooks, query SIEM logs

Banks adopt open-weight models to keep customer data resident and to control audit logging. The playbook filters to commercially licensed models and, when customer PII is present, to on-premises deployment.

Frequently asked

Can a bank run an open-source LLM on-premises?

Yes. Open-weight models such as Qwen2.5-32B or Llama 3.3 70B run on two to four data-centre GPUs behind an inference server like vLLM, so customer data never leaves the bank's network.

Which open LLM license is safest for banking use?

Apache-2.0 models carry no usage thresholds or attribution requirements. Llama community licenses permit commercial use but include a large-user threshold worth a legal review.

Start the banking assessment