Classify equipment-loan documents before the credit workflow runs

A commercial borrower uploads a bank statement and a revised equipment quote after the credit team has already opened the file. A general lending assistant may reread the entire application and regenerate its memo. A more focused intake layer asks what changed and which verification task needs the new document.

What this means for your business

For a commercial lender, document classification could reduce avoidable handoffs and repeated model work during application preparation. The objective is a cleaner package for analysts. The routing model does not make credit decisions, determine pricing or replace identity and compliance checks.

CLM-v0.1-8B scores supplied choices instead of writing an answer. Its two projection heads use a frozen Qwen3-8B encoder. That makes it a candidate for deciding which workflow should run next. It does not extract a complete record, execute a tool or prove that a document is correct. The model card and implementation are linked below.

A specific workflow: financing a CNC machining centre

A fictional machine shop requests financing for a CNC machining centre. Its application contains a dealer quote, three months of bank statements and an ownership chart. A later email supplies a replacement quote with a different spindle option and adds a bank statement that partially overlaps the earlier period.

Local parsers capture document text and page references. CLM ranks document-purpose routes: equipment-and-vendor verification, cash-flow document reconciliation, ownership-document review, application correspondence, or manual review. Applicant identity and file access are established by the lending system. Mixed packets are split before routing.

The equipment route compares the replacement quote with the existing quote, extracts the changed model specification and price, and asks the analyst to confirm the financed asset. The statement route checks account and period identifiers for overlap, then sends only new transactions to the established reconciliation process. Arithmetic, duplicate detection and required-field checks remain deterministic where possible.

The credit-memo model runs after verified changes have been assembled, rather than on every upload. CLM helps avoid unrelated specialist calls and repeated memo drafting. It cannot certify that a statement is authentic or that an applicant is eligible. Documents it cannot place confidently remain visible in an analyst queue.

A compact on-site deployment

The lightweight part is the scoring layer. The full service still needs the Qwen3-8B encoder. As a sizing calculation, eight billion weights at two bytes each occupy roughly 16 GB before runtime memory, activations and serving overhead. A single Linux workstation with a supported 24–32 GB GPU is a reasonable starting configuration to evaluate for short inputs and modest concurrency, not a tested minimum or a throughput promise.

The reference implementation pairs a vLLM pooling service with the CLM API. Keep decision inputs short, precompute reusable candidate embeddings, and batch arrivals within an acceptable waiting window. The repository's example caps inputs at 2,048 tokens; split longer documents into meaningful sections and retain page references instead of silently truncating evidence. One encoder can serve several department-specific sets of choices.

Four-bit weights would be roughly 4 GB in raw weight storage by arithmetic, but that is not the total memory requirement. Quantization changes embeddings and may change rankings. Validate the exact encoder, pooling method, precision and heads together before buying smaller hardware. The supplied heads cannot simply be attached to any small embedding model. A tiny CPU appliance or a Mac mini is not a verified drop-in deployment for this reference stack.

Put the router beside the loan-origination records

For a lender able to operate a local service, on-site hosting can keep borrower statements, ownership information and document embeddings within its controlled environment. The router should have access only to the current application and approved document store. Run any OCR and memo generation locally too if the aim is to keep inference data inside that boundary.

Maintain a record of input document version, proposed route, model configuration and human disposition. Avoid sensitive document contents in broad infrastructure logs. Local hosting supports operational control; it does not on its own satisfy every regulatory or contractual requirement. Plan for patching, disaster recovery and a manual queue before making the service part of daily intake.

Evaluate document handling, not lending outcomes

Replay historical document arrivals with analyst-approved routes, including stale quotes, overlapping statements and unclear ownership diagrams. Measure time to a verified update packet, duplicate specialist calls and missed or misrouted documents. Keep approval and pricing outcomes outside the classifier's scope. A faster route is valuable only if the credit team receives complete, correctly linked evidence.

The release is an early candidate-ranking model. Published agent benchmarks do not establish accuracy or speed on this workflow. Scores are relative to the alternatives supplied, not calibrated probabilities that a business decision is correct. Include an explicit manual-review choice, then set abstention rules using held-out examples; adding that choice alone does not make the system safe.