# distil labs releases an invoice-decision model that approves or holds vendor invoices

> A 4B Qwen3.5 fine-tune returning one of five decisions. The card reports 98 of 100 correct on synthetic test invoices.

Published 2026-09-21

Canonical: https://shipped.anth.us/decisions/distil-labs-qwen3-5-4b-invoice-decision/

distil labs published distil-qwen3.5-4b-invoice-decision on Hugging Face on 21 September 2026, by repository creation date. The card describes a fine-tuned Qwen3.5-4B that checks vendor invoices against the purchase order and goods receipt. It reasons through four checks and returns one of five decisions: approve, hold_no_po, hold_quantity, hold_price or hold_total.

It has 4B parameters in BF16 and is licensed Apache-2.0. Training used 40 seed examples plus 4,156 synthetic examples generated by a GLM 5.3 teacher with reasoning, then LoRA fine-tuning at rank 64 for four epochs, merged.

On 100 test invoices the card reports 98 correct decisions for this model. It lists GPT-5.6 Luna with reasoning high at 100, the GLM 5.3 teacher at 96, GPT-5.6 Luna with reasoning off at 81, Jev averaged over three setups at 79 and Gemini 3.5 Flash Lite at 76.

The card states the limits. The data is synthetic, English, single-currency and based on a fictional company. It is trained for one task and one policy and is, in its words, "not a general assistant." The exact system prompt and thinking mode must be used.

A small model matching a large one on a task it was built for. The test set is synthetic too.

## Sources
- https://huggingface.co/distil-labs/distil-qwen3.5-4b-invoice-decision

## Editor's first pass (AI-assisted)
- Real 4: Apache-2.0 weights with a GGUF build and a pipeline repo.
- Useful 3: One task and one policy; trained on synthetic data from a fictional company.
- Notable 2.5: A concrete example of a task-specific decision model.
- Novel 3: Distilled from a teacher model for one narrow check.