No API Keys, No Cloud Bills: Running a Document Pipeline Entirely On-Device

Описание и применение

В статье описывается локальная обработка документов с помощью модели Mano‑P на Mac mini, без API‑ключей и облачных вызовов, что обеспечивает соблюдение конфиденциальности и экономию средств.

Задача

Clients with sensitive due‑diligence documents cannot upload data to cloud vision APIs due to compliance restrictions.

Результат

On a Mac mini with 32 GB RAM, the 4B quantized model achieved a 56 % pass rate on 100 real GUI tasks, outperforming the cloud Qwen3‑VL‑Plus (39 % pass rate) and reducing average step time from 10.2 s to 7.9 s. The local setup eliminated network traffic, cut compliance review time from two months to two days, and saved roughly $800 per batch of 200 contracts.

Рабочий процесс

• 1. User drops a task to process a batch of due diligence files.
• 2. Agent searches local filesystem for the target archive and finds it in nested folders.
• 3. Agent retrieves the password from a local note and unzips the archive.
• 4. Agent reads contents, including scanned images, using a 4B vision model.
• 5. Agent performs batch extraction on each contract, maintaining a checklist to track progress.
• 6. Results are stored locally and audited via a local log file.

Доказательства

We run a document processing pipeline on a Mac mini. No API keys. No cloud inference. No network calls at all during execution.

Источник

The agent handles it step by step, all on the local machine.

Источник

Local 4B thinking model: 56% pass rate, average 7.9 seconds per step. Cloud Qwen3‑VL‑Plus: 39% pass rate, average 10.2 seconds per step.

Источник

Zero outbound connections during the entire processing run. The audit log is a local file they can inspect.

Источник

Mac mini, M4 chip, 32GB RAM. Total hardware cost under $1500. The alternative was a cloud vision API that would have billed us per image per page per document.

Источник