Which backend should I pick for offline versus cloud use?

Use LayaOnDevice for offline Core ML on the Neural Engine with no key, LayaFoundationModels for local laya-serve over HTTP, and JevFoundationModels for hosted Jev with retries and routing. Package traits let you include only the backends you ship.

How do my Generable types map to decision primitives?

Bool maps to noul binary judgment, enum or String maps to choice, Guide description maps to instructions, Guide range maps to score, and metadata exposes confidence and probabilities. Evaluation is a single feed-forward pass rather than prompt-and-parse.

What is the safe way to ship a cloud key in an iOS app?

Never hardcode TYPESAFE_API_KEY in the app binary since it can be extracted. Keep keys server-side or in CLI tools from environment variables, or route mobile traffic through ProxyTransport with App Attest or App Check.

Do I get uncertainty signals or just a bare answer?

It reports calibrated probabilities and confidence with Apple Intelligence API compatibility, and cloud clients add retry policy, cooperative cancellation and confidence routing. On-device runs need no network calls.

What Apple platforms and toolchain does it need?

Swift 6, Xcode 27 or newer, and iOS 27 or macOS 27 or newer, added as a package from version 0.3.0. It is Apache-2.0 licensed and evaluates typed Generable structs and enums in roughly 15 to 150 ms.

What is the multimodal backend it offers?

ClefFoundationModels runs the open-weight Clef (27B) and Clef-Flash (9B) models through Cloudflare Workers AI edge, Cloudflare AI Gateway, or a local runner, judging camera frames and image attachments alongside text in a single feed-forward pass.

What does the OpenAI backend offer?

OpenAIFoundationModels calls the OpenAI Decisions API at https://api.openai.com/v1/decisions with gpt-6-luna, about 150 ms latency, $0.10 per million input tokens and free output, plus multi-tenant header isolation for OpenAI-Organization and OpenAI-Project.

CommunityJev Integrations & Adapters

Jev Foundation Models

Native Swift 6 bridge that evaluates typed Generable structs and enums against app state through on-device Core ML, self-hosted HTTP, or hosted Jev endpoints.

See highlights

Open github.com

Stars: 74
Language: Swift
License: APACHE-2.0
Published by: Community
Author: peterfriese
Last verified: Oct 2026

peterfriese/jev-foundation-models

User repository on GitHub

View on GitHub

A lightweight, native Swift 6 bridge integrating TypeSafe AI's Jev System One decision model into Apple's Foundation Models framework.

Forks: 3
Last pushed: Updated Oct 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

A lightweight, native Swift 6 bridge integrating System One decision models into Apple's Foundation Models framework (LanguageModel, LanguageModelExecutor, @Generable). Evaluate strongly typed @Generable structs and enums against application state in 15–150ms with zero hallucinations, calibrated probabilities, and full Apple Intelligence API compatibility across:

Quoted from the project's README. Read the full README.

Tagscommunity jevbest catalog-import awesomejev

Works forJev for Classification

Used inJev for Software & DevTools

Languages: Swift

Highlights

On-device Laya models run locally on Neural Engine and GPU with zero network calls.
Self-hosted laya-serve transport speaks Jev-compatible POST /v1/systemone with local presets.
Hosted Jev client with automated retries, cancellation and confidence routing.
Modular targets and Swift package traits so you link only needed backends.

Watch out

Cloud keys must never be bundled in mobile binaries, so mobile cloud use needs your own authenticated reverse proxy with attestation, and on-device needs a compiled Core ML model. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

system-one

Provider-neutral TypeScript runtime for System One decisions: write choice, noul, and score code once, then run it on TypeSafe Jev, local Reflex, any compatible /v1/systemone endpoint, or a custom provider.