Docusign's tiny 900M-parameter model beats generic VLMs at tables

Docusign says $2 trillion in negotiated contract value sits unused because nobody can query what a company actually agreed to. Its fix, built with NVIDIA, wasn't a bigger model — it was a purpose-built extractor roughly 900 million parameters in size, aimed specifically at the tables generic VLMs mangle.

• The stakes — Docusign has 1.9 million paying customers, a billion users, and processes about a million agreements a day.
• $2 trillion figure — A Deloitte study cited by Docusign found $2 trillion of negotiated value locked in agreements that organizations never revisit, due to manual reading and disconnected systems.
• Agreements aren't flat — Agreements are hierarchical — one governs another, which amends a third — so answering a simple question can mean traversing 20 years of business history.
• Where extraction fails — Generic document tools and VLMs read text line by line, which breaks merged cells and nested columns in pricing tiers, SKUs, SLAs and rate cards.
• NVIDIA Nemotron Retriever — NVIDIA's Nemotron Retriever initiative builds open-source embedding, reranking, and document extraction models, and has led leaderboards including ViDoRe V1-V3 and MTEB/MMTEB.
• Nemotron Parse model — The jointly built Nemotron Parse model is a ~850-900 million parameter vision-language model designed as a single-pass extractor, not a generator, replacing separate layout and table models.
• Benchmark — NVIDIA compared Nemotron Parse against open-source models on a table-extraction benchmark referred to as 'RD table bench' and reported it improving on industry accuracy-performance tradeoffs.
• Speed gain — Docusign found Nemotron Parse extracted tables 20x faster than other open-source models tested, critical at million-plus daily document scale.
• Demo — In Agreement Manager, uploading an order form triggers automatic extraction of key terms and pricing tables, downloadable as CSV or accessible via API.
• Lessons learned — Docusign's takeaways: purpose-built models beat generic ones for specific jobs, smaller context length cuts latency and cost, and speed matters at billion-user scale.
• Hybrid pipeline — Docusign still runs OCR alongside the layout/table model in a two-route pipeline for text and metadata extraction, detailed in a companion blog post.
• What's next — The partnership plans to extend into Nemotron Retriever for document search and the NVIDIA agent toolkit, and to move the model from FP16 toward FP8 and NVFP4 with multi-token generation for further speedups.

In their words

there's $2 trillion captured in this agreement negotiated value that no one capitalizes no one goes back and gets that um data back right2:01

traditional document extraction tools or a generic VM BLM completely fail here like you know we've tried we've definitely done this because they're reading text line by line which breaks a lot of that concept within the table.3:58

This is a very small tiny C radio VLM. It's about 850 900 million parameter model.7:05

Neotron was 20x faster which helps us when we're talking about the millions and billions of scale that we're kind of serving for all of our customers.12:01

Disclosure · Both speakers work for the companies whose product and model they are presenting — Docusign's Agreement Manager and NVIDIA's Nemotron Parse — making this a joint promotional talk for that partnership.

One thing to add — One thing to add — the $2 trillion figure and the 20x speed claim both come from Docusign/Deloitte and internal benchmarking respectively, with no independent verification offered in the talk, so treat them as vendor-reported figures. The "RD table bench" name is likely a mangled auto-caption of a public table-extraction benchmark like OmniDocBench, but the speakers never spelled it out clearly enough to confirm.</note> </invoke>

One thing to try tonight
Pull one dense pricing table or rate card from a signed PDF contract and run it through a generic OCR/VLM extractor versus a table-specific one to see firsthand where merged cells and nested columns break.