Some clients in finance operate under a hard constraint: contract documents never leave their machines. Not encrypted-in-transit, not processed-by-a-trusted-vendor. Never leave. Cloud APIs are off the table no matter how good the benchmark numbers look, because once the data crosses the network boundary, the compliance sign-off process gets very long or just stops.
We recently walked through what it takes to batch-process around 200 financial contracts under that constraint, end to end. Here is the workflow, with real numbers from the README so you can sanity-check the throughput yourself.
The setup
Hardware side it is less exotic than people assume. An Apple M4 Mac with 32GB RAM covers it, a Mac mini works fine. If you would rather not dedicate a machine, Mano-P also runs from a compute stick over USB 4.0.
The stack is Mano-P, an Apache 2.0 GUI agent that runs fully local. The model open to users is the 4B one, Mano-CUA-4B-Thinking-1.1, with MLX 8-bit builds on HuggingFace. Install goes through brew:
brew tap Mininglamp-AI/tap && brew install mano-cua
First-time local setup is mano-cua check, then install-sdk, then install-model. After that every run accepts a --local flag, and in local mode screenshots and task descriptions stay on the device. The full client source is on GitHub and Apache 2.0, so security teams can audit the network paths themselves instead of taking anyone's word for it.
The per-contract loop
Each contract goes through the same loop:
- Document parsed locally into screen content the agent can see
- The 4B model reasons over it on-device, driving the GUI like a person would, no API integration with the document system needed
- Extracted fields land in the local business database
- Every step gets written to a local audit log, with what was extracted and at which screen
That last step matters more than it sounds. Because everything happened on the machine, the audit trail is complete by construction. Nobody has to reconcile log formats with a cloud vendor, and the logs themselves never turn into a new exfiltration path.
Throughput numbers
From the README performance data: the 4B model decodes at roughly 80 tokens/s on an Apple M5 Pro. On their 100-task real-machine macOS GUI benchmark, Mano-CUA-Thinking-4B passed 56.0% of tasks at 7.9 seconds per step. For comparison, a cloud general VL model, Qwen3-VL-Plus, scored 39.0% on the same suite. Adding Bash tool access, Mano-CUA 1.1 reaches 90.0%.
Per-contract time lands in the low single-digit minutes depending on layout complexity. Two hundred contracts overnight on one machine is realistic. Batch workloads do not care about latency, they care about unattended reliability, and the 56% pass rate means you should plan a review queue for failures rather than expecting lights-out operation.
The offline test
The cleanest verification we have seen for the zero-exfiltration claim: kill the network entirely and rerun the workflow. With WiFi off, the local model keeps executing tasks normally, fields extracted exactly as before. A pipeline with any cloud dependency fails this test at the first step. It is a one-line test your compliance team can run themselves, which is worth more than a vendor questionnaire.
Where this does not apply
On-device is the right call when data boundary is the binding constraint. When it is not, cloud models still win on raw capability, the 72B Mano variant that tops OSWorld specialized models at 58.2% is benchmark-only, not what users run. Match the architecture to the constraint.
If you are working on local-first document processing, the project is at https://github.com/Mininglamp-AI/Mano-P , Apache 2.0. A star helps others find it, and the issues section is open if you want to compare notes on on-device pipelines.
Top comments (0)