Sections08
Reading time9 min
EvidenceTeam guide

Comparison Summary

Local AI runs the extraction model on your own computer through Ollama. Cloud AI sends document content to a hosted model provider. In Dynamite Docs both routes return fields and table rows to the same review workspace, so the choice comes down to where inference may run, your hardware and your plan.

In Dynamite Docs, local AI and cloud AI differ in where model inference runs, not in how results are reviewed. With cloud AI, the workspace sends document content to a hosted model provider chosen by Dynamite Docs routing, or to your own provider account when you connect a key. With local AI, the browser sends the document to the Windows companion, which listens on loopback port 8756 and forwards the request to Ollama at 127.0.0.1:11434 on the same computer. The companion validates the model response before it returns fields, rows and confidence signals to the workspace. On both routes, sign-in, short-lived token issuance, plan state and the review workspace stay hosted, so the computer still needs internet access. File storage is a separate choice: browser-only storage keeps the original out of the cloud library on either route. Local inference is available on Starter, Pro and Scale. Hosted and own-key routes are available on every plan.

Neither route makes the result correct by itself. Model choice, scan quality and table density affect every route, so keep source review and approval in the workflow whichever one you pick.

This page also covers Ollama vs cloud LLMs. Ollama is the local runtime the companion uses; cloud LLMs are the hosted or own-key models in the provider registry.

Test a document

Run the same file through a hosted model first, then compare it with a local model.

Local AI
Ollama on your Windows PC through the companion
Cloud AI
Hosted routing or your own provider key
Still hosted
Sign-in, tokens, plan state and workspace
Local plans
Starter, Pro and Scale

Fit register

Which route fits your documents

Start from the rule your documents must follow, then check whether your hardware and plan support that route.

Choose local AI when model calls must stay on your machine

Use the Windows companion with Ollama when policy says document content must not reach an external model API. You need Starter, Pro and Scale, a vision model for scans and photos, and hardware that can run it within the companion timeout.

Choose hosted cloud AI when you do not want to manage models

Hosted routing works on every plan without installing anything. Document content goes to the selected provider, so check that provider terms, training rules and residency fit the file before you upload it.

Choose your own provider key when the cloud account must be yours

Bring your own key when the provider contract and billing should sit under your account. Policy controls such as no-training-only routing can narrow the providers a sensitive document may use.

How we maintain this comparison

Written and maintained by the Dynamite Docs team. This page compares two processing routes inside the same product, not two vendors.

Allowances and limits come from the current plan table and companion documentation. We have not run a published head-to-head benchmark between local and cloud models on a shared corpus.

Test the same files on both routes before choosing. A model that wins on one invoice layout may lose on a dense scan or a long table.

Decision table

Local AI and cloud AI side by side

The table compares the three model routes in Dynamite Docs. Review, correction and export work the same way on each.

Swipe horizontally to compare every column

CriterionLocal AI with OllamaHosted cloud AIYour own provider key
Where inference runsYour Windows PC through OllamaProvider selected by Dynamite Docs routingProvider account you connect
Document content leaves the PC for inferenceNo, the request stays on the loopback pathYes, to the selected providerYes, to your provider
Internet access neededYes, for sign-in, tokens and the workspaceYesYes
Plan accessStarter, Pro and ScaleEvery planEvery plan
SetupInstall Ollama, a vision model and the Windows companionNoneAdd and verify an API key in AI Settings
Speed and capacity set byYour CPU, GPU memory and model sizeProvider capacity and plan limitsProvider limits on your account
Request limits150 MB upload, 120-second default timeoutPlan page and file limitsPlan limits plus provider limits
Dynamite Docs page equivalents (PE)1 PE per successfully processed page, same as hosted extraction1 PE per successfully processed page on hosted models0 PE on Pro and Scale when BYOK is unlimited; provider bills separately
Review and exportSame workspace, Excel, CSV or JSONSame workspace, Excel, CSV or JSONSame workspace, Excel, CSV or JSON
Test a document

Run the same file through a hosted model first, then compare it with a local model.

Working procedure

How to compare local and cloud AI on your own files

Dynamite Docs logoDynamite Docs
The handoff
Source document: Local AI vs cloud AIPick representative documents. Run a hosted model first. Run the same files locally. Decide per document type. Not extracted customer data.THE LOCAL INFERENCE PATHBrowser workspaceReview the documentSigned local connectionLoopback companionOllama on your computerRun local inferenceSamesource-linked…Local inference requires the companion and a supported local model.
  1. 01Pick representative documents
  2. 02Run a hosted model first
  3. 03Run the same files locally
  4. 04Decide per document type
Keep the original beside the result.
  1. 01

    Pick representative documents

    Include a clean PDF, your weakest scan and your longest multi-page table. Write down the fields and row counts you expect from each.

  2. 02

    Run a hosted model first

    Extract each file on the hosted route and record missing fields, wrong amounts and broken rows after review.

  3. 03

    Run the same files locally

    Install Ollama and a vision model, start the companion, choose Local in AI & Processing and extract the same files. Check ollama ps to see whether the model is using GPU memory.

  4. 04

    Decide per document type

    Compare corrections and run time per file. Keep local inference for documents that must stay on the machine and use the cloud route where your policy allows it.

Dynamite Docs logoDynamite Docs
The review
Source document: Local AI vs cloud AISign-in, short-lived local tokens, plan checks and the web workspace remain hosted services.. Files saved to the cloud library are stored by Dynamite Docs whichever model route you use. Choose browser-only storage to keep the original on your computer.. Local model quality varies with the model, quantization, document type and hardware. A smaller model may miss small text or dense tables.. Not extracted customer data.SOURCE DOCUMENTSource document🔍 Click to zoomReview before exportLocal AIOllama on your WindowsPC through the…Compare with the source on the leftAccept after reviewKeep the original beside theresult.Original stays…
Local AI
Ollama on your Windows PC through the companion
Cloud AI
Hosted routing or your own provider key
Still hosted
Sign-in, tokens, plan state and workspace
Keep the original beside the result.

Provider independence

Switch routes without changing the review workflow

The model route is a setting, not a separate product. Fields, rows, confidence signals and corrections appear in the same workspace whichever route produced them.

Local-only routing can stop a document from falling back to an external provider. With that rule active, an unavailable companion, Ollama service or model stops the extraction instead of sending it to the cloud.

Compare every AI route

Local

Ollama runs the model on your Windows PC.

Hosted

Dynamite Docs routes the request to a hosted provider.

Own key

Your provider account runs and bills the request.

What local AI does not change

Local inference moves the model call to your computer. It does not make the whole product self-hosted or offline.

  • 01Sign-in, short-lived local tokens, plan checks and the web workspace remain hosted services.
  • 02Files saved to the cloud library are stored by Dynamite Docs whichever model route you use. Choose browser-only storage to keep the original on your computer.
  • 03Local model quality varies with the model, quantization, document type and hardware. A smaller model may miss small text or dense tables.
  • 04Every extracted value still needs review before it reaches your books, a filing or another system.

Sources and review date

Questions

Check the edge cases before you buy

Is local AI more private than cloud AI?+

For model inference, yes: on the local route the document goes from the browser to the companion and Ollama on the same computer instead of an external model API. Account, token and workspace services remain hosted, so review the full storage and retention path in the private extraction guide.

Ollama vs cloud LLMs: which should I use for document extraction?+

Use Ollama when document content must not reach an external model provider and your PC can run a suitable vision model. Use a cloud LLM when you want a hosted model without installing or managing one. Test both on the same files before deciding.

Does local AI work offline?+

Ollama runs the model on your PC. The web workspace, sign-in and account services require internet access.

Which local models work for scanned documents?+

Scans and photos need a model that accepts images. The local AI guide starts with gemma3:4b and explains how to compare candidate models on your own files.

Can I use both routes in the same account?+

Yes. On Starter, Pro and Scale you can keep a provider key and a local model configured in AI & Processing and pick the route before you extract. Local-only routing rules can restrict sensitive documents to the local model.

Do local Ollama extractions use page equivalents?+

Yes. A successfully processed page costs 1 PE on local Ollama, the same as a hosted extraction. Failed pages cost 0 PE. Pro and Scale own-key runs do not consume PE, but that rule applies to BYOK cloud models, not the local Ollama route.

Why Switch

Why finance and operations teams switch to Dynamite Docs.

Self-Serve in 60s

No mandatory sales calls or multi-week pilots. Test 25 pages free immediately on your real documents.

Provider Independence

Bring your own OpenAI, Anthropic, Gemini, or local Ollama keys with zero markup and unlimited runs on Pro.

Multi-Page Tables

Handles borderless tables, multi-page line items, and messy layouts without rigid pre-trained templates.

Clean Native Exports

Export straight to Google Sheets tabs, Excel (.xlsx with typed cells), CSV, or stream to webhooks.

First document

Test the file that makes the current process slow.

Test a document

Run the same file through a hosted model first, then compare it with a local model.

Can we save optional cookies?

Essential cookies keep you signed in. Optional analytics and masked session replays help diagnose errors. Change your choice in Cookie settings. Privacy policy.