Model families

You can select between 3 model families: Spark, Flux and Nova. A fourth family, Pulsar, isn't released yet and will be available from version 4.

The families read the same document types and return the same fields. They differ in how much compute runs on every page, and in how hard a document can get before extraction starts to slip.

Every feature is supported on every family. Only the model underneath changes.

The families

Pulsar isn't released yet. It will be available from version 4.

How accuracy moves

Two things raise accuracy independently. Each new version of a family is better than the one before it, and within any single version a larger family scores higher than a smaller one.

The chart is illustrative. It shows the ordering, and the axes carry no measured values. Three things follow from it. Each version gains less than the one before it, as accuracy runs into a ceiling. The gaps between families narrow near the top for the same reason. Moving up a family still gains more than waiting for the next version of the family you are on.

Pulsar has a single point because the family starts at version 4.

Agentic extraction on Pulsar

Pulsar models are large enough to call tools while they extract. Instead of reading a page once and returning fields, a Pulsar run can look again. It can re-read a region, check a value against another page, run a calculation, or query a lookup, and it keeps working on the parts of a document it isn't confident about.

This is called test time compute. The model does more work on the documents that need it, instead of the same fixed work on every page. It raises accuracy on ambiguous and cross-referenced documents. It is also why Pulsar is slower and costs more per page than Nova.

Spark, Flux and Nova read a page in a single pass. They're faster and cheaper because the work per page is fixed.

Versions

Each family ships new versions over time, like Spark OCR 2 and Spark OCR 3. The family sets the price. The version sets the model.

spark-ocr-3
└────┘ └─┘ └┘
family  |  major version
     capability

Credits per page stay roughly the same across versions within a family. A new version doesn't put you in a new price bracket. Price is set by the family.

Older versions stay available. To move to a newer version, change the version on your model. The family, your fields and the output format don't change.

Pulsar starts at version 4. There is no Pulsar OCR 1, 2 or 3. The family arrives with the version 4 release.

Why the families cost different amounts

Each family runs a different amount of compute on every page, and the credits per page reflect that compute. They don't reflect the type of document or the number of fields you extract.

Pulsar sits above the others because the amount of compute it runs varies. It spends extra passes on the pages it's least sure about.

Larger families resolve pages that smaller ones send to review: a smudged scan, a table with no ruling lines, a handwritten amendment, a field that only makes sense given page 14.

On a clean, well-structured document all the families will usually agree. The difference shows up on the hard pages. If your documents don't have hard pages, a larger family won't extract any better, it'll just cost more.

Extraction is charged per page, not per document. A 10 page PDF run through an extraction block is 10 runs. See Consumption based Billing for how runs are counted on every other block type.

Which family should I use?

Start with Flux. It clears most document sets.

Move down to Spark if your documents are clean and your volume is high. Move up to Nova if your review queue is the expensive part and the pages are hard. Pulsar, once it's available, is for documents where the answer depends on more than one page.

Test on your own documents before you decide. Take 200 of your files, including the messy ones, run each available family against your fields, and compare field level accuracy on the worst tenth. That is where the families differ, and where your review cost comes from.

You can set a different family on each workflow, so each document type runs on the family it needs.

FAQ

1. Does a bigger family always extract more accurately?

   No. On clean, well-structured documents the families largely agree. The difference shows up on degraded, unusual or ambiguous pages.

2. Do credits per page change between versions?

   Not meaningfully. Credits per page stay roughly the same across versions in a family. The family sets the price, not the version.

3. What happens when a new version ships?

   Nothing, until you switch to it. Older versions stay available.

4. When can I use Pulsar?

   Pulsar isn't released yet. It will be available from version 4. Talk to your account manager if you want to be told when it ships.

5. Can I use more than one family?

   Yes. Set the family per workflow, so each document type runs on the family it needs.

6. Do I get charged for retries if extraction fails?

   No. If a block fails and needs to retry, you're only charged for the successful run.

7. How do I estimate my monthly cost?

   Multiply the family's credits per page by your monthly page volume.

How to configure

Model families aren't selectable in workflow setup yet. To run a family other than the default, contact your account manager and we'll set it up on your account.