Models

Our models.Small, focused and built to run.

We fine-tune and train small models for specific jobs: two voice models that run on the edge, a working memory for AI workflows, and parsing models that read any document.

Speech enhancement · edge

Hush

Removes noise and keeps the voice, on the device.

Learn more
9:41●●● ▮
FSField support02:14
Hush
Noise removal on device
Microphone input
Street, wind and traffic
Hush output
Voice keptOn device
Illustrative interface
Turn detection · edge

Cue

Knows when a speaker has finished, so agents reply at the right moment.

Learn more
9:41●●● ▮
Voice agent · listening

I need to move my booking to Thursday, and…

…also add one more person.

Done. Thursday at 10, for two people.

Cue · turn detection
Person
Agent
Pause · still talkingEnd of turn
Running onThis deviceNo round trip to a server for turn decisions.
Illustrative interface
Working memory · Python

Thredmark

Source-backed working memory for evolving AI workflows.

Learn more
Thredmark · deal-review workflowSL
Working memory
142 itemsPython API
FactSourceVersionUpdated
Renewal is annual, 30 days noticecontract.pdf · p4v22h ago
Customer prefers weekly deliverycall notes · 12 Marv13d ago
QA bar raised to 95% agreementspec-v3.md · §2v31h ago
Pilot covers oncology onlyemail · 9 Marv15d ago
Data must stay in EU regiondpa.pdf · §7v16d ago
contract.pdf · page 4

…this Agreement renews annually unless either party gives 30 days written notice before the renewal date…

Evaluation run
RecallGrounded in sourceStale facts replaced
Illustrative interface
Document parsing · small models

Parsing models

Reads documents in any format at 99% accuracy, faster and cheaper.

Learn more
Parsing · batch 0318SL
INVOICENo. 4471
Harbor & Pine Co.14 Mar 2026
142.46
182.28
164.61
444.10
199.56
195.14
TotalUSD 1,842.50
Extracted fieldsPassed
IssuerHarbor & Pine Co.
Date2026-03-14
Total1842.50
CurrencyUSD
Line items6
Formats in this batch
PDFScanXLSXPhoto
99% accuracy
Illustrative interface

Why small models

A focused model can be faster, cheaper and easier to run than a large general one, and it can run where the data is: on a phone, a headset or a server you control.

Each model is trained and evaluated with the same expert data and benchmarks we deliver to customers.

Questions, answered.

Does Saolabs build its own AI models?

Saolabs fine-tunes and trains small, focused models: Hush, Cue, Thredmark and its parsing models. It does not build a general-purpose foundation model.

Can we license a model?

Talk to us about your use case. Access and terms are agreed per engagement.

Every safe model
has an expert behind it.