OpenJevSwift

A native Swift implementation of OpenJev, the open, Jev-compatible "System One" decision server. Send it a state and typed questions (noul, choice, score); it reads every answer from a model's probabilities rather than generating text, so an answer cannot go off-schema: a noul's is the probability of yes, and a choice's or a score's the probability of every option, with a confidence. One openjev binary serves DiffusionGemma 26B-A4B or JevK5 on MLX, or Verdict or Laya on Core ML, on a Mac, with upstream's exact wire API, so TypeSafe's SDKs work against it unchanged; the same libraries answer requests inside an iPhone or Mac app.

OpenJevSwift is an independent project. It is not affiliated with or endorsed by TypeSafe AI (the makers of Jev), by Google DeepMind or NVIDIA (DiffusionGemma), or by the authors of the other models it serves. Jev, TypeSafe, Gemma and other names are the property of their respective owners.

Quick start

On an Apple silicon Mac with macOS 15 or later, Xcode 27 and its Metal Toolchain component (docs/development.md has the one-time install):

git clone https://github.com/Algorythm-Canada/OpenJevSwift.git
cd OpenJevSwift
swift build -c release --product openjev
.build/release/openjev serve --backend verdict

The first start downloads Verdict's converted Core ML package, tokenizer and calibrator (about 310 MB) into Application Support and checks each file's SHA-256; the server listens on 127.0.0.1:8080 once the model has loaded and warmed up. From another terminal:

curl -s localhost:8080/v1/systemone -H 'content-type: application/json' \
    -d '{"model":"jev-latest","state":"The deploy failed twice and the site is down.","questions":{"urgent":{"type":"noul","instructions":"Is this urgent?"}}}'

{"model":"verdict-1.4","answers":{"urgent":{"type":"noul","noul":0.5910340547561646}},"usage":{"input_tokens":45,"output_tokens":0}}

openjev decide answers one request without a server and prints the bytes the server sends:

echo '{"model":"jev-latest","state":"The deploy failed twice and the site is down.","questions":{"urgent":{"type":"noul","instructions":"Is this urgent?"}}}' | .build/release/openjev decide --backend verdict

Ctrl-C stops the server gracefully. docs/deployment.md covers the settings, a launchd job, the logs and the exit statuses, and the DocC article "Making decisions in an app" loads the same models inside an app.

Requirements

Backend	Model	Mac	Memory
verdict	verdict-1.4, 151M parameters, Core ML	Apple silicon, macOS 15 or later	1.6 GB with the functions one-question reads load, 2.8 GB with all six
laya	laya-1.0, 421M parameters, Core ML	Apple silicon, macOS 15 or later	4.7 GB with the functions one-question reads load, 8.9 GB with all eight and up to 9.7 GB at peak; with OPENJEV_ENCODER_FUNCTIONS=2, 2.1 GB for one-question reads and up to 4.4 GB at peak
mlx	openjev-0.1, DiffusionGemma 26B-A4B, 4-bit, MLX	Apple silicon	about 16 GB to load, 17.3 GiB in service with short prompts and up to about 3.6 GB more for cached long prompts; 32 GB or more recommended
jevk5	jevk5-0.2, JevK5 (Qwen3.5-4B), 8-bit, MLX	Apple silicon	6.0 GB once loaded, up to 11.0 GB in service with OPENJEV_MLX_CACHE_LIMIT_GB=4; 3.6 and 8.9 GB with the 4-bit conversion

Building needs Xcode 26.4 or later. The mlx backend also needs MLX's Metal shaders, which Swift Build compiles with the Metal Toolchain: Swift Build is the default with Xcode 27, and Xcode 26 takes --build-system swiftbuild. In an app, OpenJevCore runs on macOS 14 and iOS 17 or later, the Verdict and Laya backends on macOS 15 and iOS 18 or later, and JevK5 on Apple silicon (it builds for iOS; it has not run on an iPhone yet). Linux builds the core, the server and the openjev tool for the tests, without any backend.

Status

Milestones 0 to 3 are complete, every work issue in them closed: the foundations, the decision engine core, DiffusionGemma reads on MLX, and the Jev-compatible HTTP server with the openjev tool. Milestone 4, the read extensions and images, is in progress: steps, samples and sequential are verified end to end on the DiffusionGemma checkpoint (#43, #44 and #45), and images remain. Verdict and Laya, from milestone 6, and from milestone 7 the JevBench comparison with upstream and DiffusionGemma's calibration report (#61 and #62), are done too. JevK5 (#55), also from milestone 6, is served, and gives its author's published top answer on 230 of JevBench's 231 items (D-052).

Not there yet:

• Images in requests: #46, #47 and #48.
• think, a thought before the read: #50, #51 and #52.
• POST /v1/chat/completions: #53.
• The CLM model: #59.
• Release 0.1.0, with a version tag a package can depend on: #65.

docs/08-implementation-plan.md has the milestones and the issue index.

Compatibility

Apart from the recorded differences, everything up to the model's probabilities is upstream's byte for byte: the prompts, the answer templates, the canvases and seeds, the request validation, the error bodies, the headers and the /v1/models listing, all checked against fixtures upstream's own code writes. The probabilities agree within measured bounds: DiffusionGemma's within decisions D-014's and D-048's on the 63 oracle reads (the top label on 91.7% of slots, and on 139 of the 140 where mlx-vlm's top two are at least 0.5 apart), and between the two servers on 333 JevBench and TypeSafe items; Verdict's and Laya's with upstream's top answer on all 666 items; JevK5's, on its 8-bit conversion, with its author's published top answer on 230 of JevBench's 231 items and the same token counts on all of them. The differences, among them a stricter JSON parser, the 503 for any backend failure and the features not built yet, each have a decision record. docs/compatibility.md has the three tables and the matrix of what runs on macOS, iOS and Linux.

Documentation

• API documentation. The DocC catalogs of OpenJevCore, OpenJevEncoders, OpenJevDiffusionGemma, OpenJevLetterReadout and OpenJevServer: getting started in an app, the request and answer types, implementing a backend, running the server and the configuration reference. The Documentation workflow builds them whenever the sources change and, once GitHub Pages is enabled for the repository, publishes them to https://algorythm-canada.github.io/OpenJevSwift/. make docs builds the same site locally (docs/development.md).
• docs/deployment.md: running openjev serve on a Mac.
• docs/compatibility.md: what is identical to upstream, within tolerance, or different.
• docs/credits.md: the models, their authors and licenses.
• docs/quality.md and docs/benchmarks.md: answer quality against upstream, and speed and memory.
• docs/README.md: the index of the design documents, from how upstream works to the decisions and the conformance strategy.

License and credits

Apache-2.0, the same as upstream OpenJev. Code ported from mlx-vlm (MIT) keeps its copyright notice in each file's header; THIRD_PARTY.md lists every referenced project at its pinned revision. The models are other people's work and keep their own licenses: docs/credits.md credits each, with where this project gets its weights. No weights are part of this repository.