Add the Clef library product to the application's dependencies. Requires Apple Silicon and macOS 15+ (MLX); leave several GB of free disk space for the approximately 5.3 GB checkpoint and adequate unified memory for inference.

import Clef

let model = try await Clef.fromPretrained()
let result = try model.decide(
    state: "Please turn the kitchen lights on.",
    questions: [
        ClefQuestion(id: "action", instructions: "Which action was requested?",
                     kind: .choice(["on": "Turn lights on", "off": "Turn lights off", "other": "Other request"])),
        ClefQuestion(id: "urgent", instructions: "Is this urgent?", kind: .noul())
    ])
for decision in result.decisions {
    print(decision.id, decision.probabilities)
}

Clef.load(from:) loads an existing local export without network access. fromPretrained(cacheDir:offlineMode:progressHandler:) uses the package's model downloader and cache. The default model is aufklarer/Clef-flash-9B-MLX-4bit. The downloader follows the repository's current revision; for reproducibility download the pinned revision listed in the model document to a dedicated directory, then use load(from:).

selectedOption and confidence describe the highest-probability option. For .noul(), read noul for the probability of true even when false wins. For .score(["normal", "urgent"]), score is the probability-weighted criterion index. The API never calls a tool.

The main CLI accepts a JSON request and downloads the published model on first use:

speech clef decide request.json
speech clef decide request.json --model-dir /path/to/Clef-flash-9B-MLX-4bit
speech clef decide request.json --offline --max-tokens 4096

Output is JSON with decisions, input_tokens, load_seconds, and decision_seconds. --offline uses only the local cache. --model-dir loads an existing directory without downloading. --max-tokens accepts 1–16384 and defaults to 4096. The library's ClefRequest decodes this same JSON format and exposes resolvedQuestions() for applications.

For the smaller standalone build:

swift build -c release --product clef-decide --disable-sandbox
scripts/build_mlx_metallib.sh release
.build/release/clef-decide /path/to/Clef-flash-9B-MLX-4bit request.json

{
  "state": "Please turn the kitchen lights on.",
  "questions": [
    {"id": "action", "type": "choice", "instructions": "Which action?", "choices": {"on": "Lights on", "off": "Lights off", "other": "Other"}},
    {"id": "urgent", "type": "noul", "instructions": "Is this urgent?"},
    {"id": "priority", "type": "score", "instructions": "Rate urgency", "levels": ["normal", "urgent"]}
  ]
}

The command emits JSON with probabilities, input token count, and decision_seconds (encoding through materialized probabilities, excluding model load). A single first inference includes warm-up and is not a steady-state latency benchmark. CLI input uses an ordered field array and a text state; it is not a drop-in SystemOne HTTP API.

swift test -c release --no-parallel --filter ClefEncodingTests
swift test -c release --no-parallel --filter E2EClefHeadParityTests
CLEF_MODEL_DIR=/path/to/model swift test -c release --no-parallel --filter E2EClefTests

Run model tests and benchmarks sequentially. The head parity fixture uses the released PyTorch implementation with deterministic small tensors, independently checking logits and schema piece boundaries. Full model tests require local model files; they do not download weights implicitly.

On M5 Pro (48 GB), one loaded model handling a repeated 303-token request reached 5.11 GiB peak RSS and 6.81 GiB peak macOS physical footprint. The latter includes unified-memory costs that RSS misses. Twenty warmed runs averaged 291 ms; model loading is excluded. Budget additional memory for longer inputs and the rest of your application. See the measured report for scope and methodology.

Timings vary with background activity: a later correctness run measured 462–897 ms while other applications were active.