Question 1

Is this a hosted Jev wrapper, or does everything run on my own Mac?

Accepted Answer

Everything runs locally. FluidUse is a Swift package that reads a running application through the macOS Accessibility API and scores decisions on Apple silicon with no cloud calls. Typed choice, score and noul answers come from open models converted to Core ML, served through FluidAudio on the Neural Engine.

Question 2

Is the laya model here the same thing as TypeSafe's hosted Jev?

Accepted Answer

No. laya is an independent open Jev-style decision model from Convai Innovations: a 322M mmBERT encoder plus decision head that answers typed choice, score and noul questions with calibrated probabilities in one pass and no generated tokens. TypeSafe's hosted Jev is a separate service this package does not call.

Question 3

Are those accuracy and speed numbers from an independent evaluation?

Accepted Answer

They are laya's own published suites, rebuilt from upstream's scripts: 3,899 questions where the Core ML buckets match the PyTorch reference accuracy per suite, including AG News 0.935 and guardrails 0.808, at 5.2 ms median per question against 61.6 ms for PyTorch on the same Mac's CPU. Treat them as a reproduction claim, not a third-party benchmark.

Question 4

What can it actually do end to end, and how big are the downloads?

Accepted Answer

Forms, mainly. AccessibilityFormDriver reads fields, labels and values from any NSRunningApplication and WebFormDriver does the same for an embedded WKWebView, then a 706K-parameter CUA-S1-FORMS specialist scores each actionable field at about 1 ms per decision. The larger models add text and image decisions at hundreds of megabytes per bucket, and GLiNER 2.5 errors on inputs over 128 tokens.

Question 5

How is the form filling actually wired together?

Accepted Answer

The Swift package exposes a CuaS1FormsManager with a 706K-parameter form specialist from Cua converted to Core ML and served by FluidAudio. AccessibilityFormDriver takes any NSRunningApplication and snapshots fields, labels and values; WebFormDriver does the same for an embedded WKWebView. FormSchema renders the options for each field, the model scores it, and then driver.type or driver.click does the typing. DocumentEntities turns a PDF or Label: value text file into the profile.

Question 6

What are the other models, and how big are the downloads?

Accepted Answer

Several beyond laya. KevFastManager runs Kev-0.8B on the GPU, a pinned checksummed snapshot of about 3.5 GB on macOS 15 and iOS 18. ClefVisionManager runs clef-vision-0.8b, a 0.92B image model distilled from Cloudflare's Clef-flash, at 377 MB for the vision tower plus roughly 960 MB per language-model bucket you load, with 92.2 percent held-out accuracy against a 94.1 percent teacher. GLiNER2Manager downloads 291 MB base or 385 MB multilingual classification packages.

Question 7

What are the known limits of the on-device path?

Accepted Answer

Several are documented. The GLiNER 2.5 packages use an L128 bucket, support up to eight labels and raise an error on inputs over 128 tokens. laya's Core ML buckets cover 128, 256, 512 and 1024 tokens with 32 options, and the largest bucket truncates the state on the right like laya's max_len. They also expose the classification path only: the native entity, relation and record extraction heads are not available through this Swift manager.

FluidUse

About this project

Highlights

Questions about this entry

Reactions & coverage

Related terms

More Jev resources like this

From the Jev guides

From the Jev community