AI Weekly Malaysia

Dharma-AI published a Hugging Face blog post for Dharma Field Extractor, a 2-billion-parameter model that takes a document image plus either an open request or an exact schema and returns structured JSON — invoice line items as ordered lists, ID fields under your own column names, medical requisition fields ready to schedule — with one checkpoint and no per-type fine-tuning. It follows their April DharmaOCR-LITE OCR model, and the post frames it as the opposite tradeoff: OCR gives you the whole page faithfully, the Field Extractor gives you only the fields a downstream system consumes. The post claims it outperforms every frontier LLM they tested at a fraction of the cost and latency, but the excerpt contains no benchmark numbers, no pricing, no license, and only 11 upvotes on the post.

Why: If you currently burn frontier-API calls to turn invoice, ID, or requisition images into typed JSON, a 2B checkpoint is plausibly cheap enough to self-host at batch volume — but this post gives you zero eval numbers, no price, and no license, so the only decision it supports today is 'put it on a test set', not 'migrate'. Build a 50–100 document eval from your own scans with your exact target schema and compare field-level accuracy and cost per document against your current prompt-based pipeline before touching production.