September 2026
The model signs nothing
We built Vulcan 2.5 to read supplier paper and published its weights. The post explains what you can check and who still has to approve the record.
Open weights
Models
One per industry
We make models that read the paper your industry runs on. Vulcan 2.5 is published; two models are in training and four are planned.
Commerce and home goods
[COM]
Published, September 2026
Vulcan 2.5
Vulcan reads the lab reports, spec sheets and vendor packets your suppliers send. It hands your team a product record with the page behind every value. If the paper doesn’t say, the box stays blank for a person to check.
- Vendor operationsLulu & Georgia · Live
- Product complianceLulu & Georgia · Live
- PLMLulu & Georgia · Live
- Vendor intakeIke Behar New York · Scoped
- Retailer item setupShakewell
Supplier records | Vulcan 2.5
A pipeline in six stations. Paper: lab reports, spec sheets and new-vendor packets arrive from suppliers. Pages: the paper stays attached to the values read from it. Model: Vulcan 2.5 reads those values. Values: each value keeps its source page. Record: the product record has filled fields and a blank where the paper gives no answer. Review: the record and its source pages reach a person, who decides whether to approve.
Dental and healthcare revenue cycle
[RCM]
In training
Altadena 1.5
We’re training Altadena to read eligibility, benefits, remittances, claim status and attachments. It will hand the front desk each patient’s plan, leaving blank whatever the payer didn’t say. A blank beats a guess. In its first test, on ten breakdown sheets it had never seen, it got 93.0% of all the boxes right, against 70.9% for OpenAI’s GPT-6 Luna.
- Insurance verificationHorizon Dental Care · Live
- Claims follow-upHorizon Dental Care · In build
- Claims and denialsArroyo Dental Group · In discovery
Dental breakdowns | Altadena 1.5
A pipeline in six stations. Evidence: a patient’s insurance data from three sources, the payer’s portal, the plan’s PDFs and the portal read again. Context: the patient’s whole file, read once. Model: Altadena 1.5, which is in training on sheets our team checked by hand. Answer heads: one per kind of box, each choosing from that box’s list of allowed answers, blank included, so it never makes up a value. Breakdown: the sheet of about 290 boxes, filled. Review: each box carries how sure the model is; the boxes below the bar are flagged verify, and a person checks them.
Accounting and payroll
[PAY]
In training
Blob 2.0
We’re training Blob to read the hours your clients send in time sheets, attached timecards and emails. It will draft a pay line for each worker, with the source beside it. A missing punch becomes a question for your team to answer.
- Payroll automationLazar & Company · Live
- Email intakeLazar & Company · Live
- Client onboarding and relationsLazar & Company · In build
Payroll drafts | Blob 2.0
The work Blob 2.0 is being trained to do, shown in six stations. Inbox: hours arrive in email, attached timecards and time sheets. Hours: the figures stay on their source pages. Model: Blob 2.0 is in training, with hours beside the model on a linked inset. Pay lines: it will prepare a line for each worker with the source beside it. Questions: a missing punch stays open and becomes a question on the line. Review: a person answers the question and decides whether to approve the draft.
Real estate
[CRE]
Planned
Ridge 1.0
Ridge is planned for the rent rolls and leases a building runs on. It will hand the asset manager each date with the page it came from. We haven’t trained it yet.
- Asset managementAmoroso Companies · In build
- Investor reportingAmoroso Companies · In build
- Deal packets10mark Holdings · Live
Lease dates | Ridge 1.0
Ridge 1.0 is planned and not trained yet. The six stations show its intended job. A rent roll and lease pages arrive, with date marks on the paper. The model will read the dates, keep each date with its source page, and hand the source-linked dates to the asset manager. The last station shows the asset manager reviewing them.
Wealth management
[WM]
Planned
Arroyo 1.0
Arroyo is planned for custodian statements and the forms that open or move accounts. It will hand the advisor a report with every figure linked to its statement. Training hasn’t started yet.
Statement figures | Arroyo 1.0
Arroyo 1.0 is planned and not trained yet. The six stations show its intended reporting job. Custodian statements arrive, their figures marked on upright pages. The model will read the figures and keep each beside its source statement. They will go into a draft quarterly report. An advisor will approve the report before anything goes out.
Legal services
[LEGAL]
Planned
Phantom 1.0
Phantom is planned to read a case’s intake notes, orders and the dates inside them. It will draft from that record, but an attorney must review it before anything becomes advice. We haven’t trained it yet.
- Client intakeThe Peoples Law Project · In discovery
- DeadlinesThe Peoples Law Project · In discovery
- Document assemblyThe Peoples Law Project · In discovery
Case paper | Phantom 1.0
Phantom 1.0 is planned and has not been trained. The six stations show its proposed job. Intake inquiries and case paper, including notices and orders, reach the model. It would gather the facts into a matter record, with screening for eligibility and conflicts, and put dates on the matter and calendar. It would draft letters and pleadings from that record. An attorney reviews and signs; nothing it drafts is advice until an attorney makes it so.
Live events and event media
[MEDIA]
Planned
Svengali 1.0
Svengali is planned to read an event’s paper from the first inquiry to the settlement. It will hand your team the holds to confirm and the settlements to check. Training hasn’t started yet.
- Booking intakeYourShot USA · In discovery
- Crew schedulingYourShot USA · In discovery
- SettlementsYourShot USA · In discovery
Event paper | Svengali 1.0
Svengali 1.0 is planned and has not been trained. The six stations show its proposed job. An inquiry from email or a web form and the event’s paper reach the model. Crew availability and skills feed it from a separate plate. It would put a hold on the calendar for a person to confirm, match the crew and draft the call sheet, then draft the settlement from the contract, add-ons and expenses after the event. A person checks every settlement before it goes out.
Open weights
With open weights, you can keep the file that read your paper.
A model learns from examples and stores what it learns in numbers called weights. With a rented model, you send a document to its owner’s servers and get an answer back. The owner can change the model or retire it without leaving you a copy.
Open weights means you can download that file and run it on your own machines. Keep a copy, and you can run the same version five years from now.
The same paper gets the same answer.
Better at less
One industry each.
A general model answers questions across subjects, from lab reports to poems. We train each of ours to read one industry’s paper. Altadena 1.5, the dental model we’re still training, got 93.0% of all the boxes right on ten breakdown sheets it had never seen. OpenAI’s GPT-6 Luna got 70.9%. We published the test.
No second opinions
Monday, Tuesday, next year.
A general model can give you two readings of the same lab report on different days. A lead test doesn’t have two results. Vulcan 2.5 gives one reading every run when you keep its version and settings fixed. Nobody wants a calculator with opinions.
The receipt names the model
Run it again.
Every write MarzyOS makes records the model call behind it and its version. If a provider changes or retires a rented model, that receipt may name something you can’t run again. Keep the published weights and you can check the reading against the same paper next year.
Your cloud or ours
Same file either way.
You can run Vulcan 2.5 in your cloud, where the paper stays with you. You can also run it in ours; either way, it reads from the same weights.
What open means here.
Vulcan’s weights are public.
Anyone can download Vulcan 2.5, run it, test it and build on it under its license. The other models are in training or planned. Publishing their weights depends on the rights to the paper they learn from.
Your data follows your rules.
Your access rules still apply when a model reads a document. MarzyOS keeps customers apart, checks permissions and records the model behind each write. We list clients above to show the work we do together. Each workflow keeps the status on its client’s page; it doesn’t tell you which model reads the paper.
Open still needs checking.
Published weights let you check what a model read; they don’t make it right. A person or a rule you wrote down must approve the result. The model signs nothing.
We still use general models.
Vulcan 2.5 is the only model we’ve released so far. General models handle the rest of the work in MarzyOS, and each receipt names the one that did the reading.
The work around the model
A model can read a page, but it still needs somewhere to put the answer. We build the map of your business, the receipts and the steps your team uses to approve the work. Published weights let you hold the part that reads.
February 2026
The agent is the easy part
Bring us the paper your team retypes.
Start
One stack of documents
Bring a stack of lab reports, benefits pages, time sheets, leases, statements or court notices. We’ll show you what gets read and the page each value came from.
Get in touch Read the launch post