FluidUse 0.3.0

GLiNER2.5-Decide on Core ML

• GLiNER2Variant.decide (128 tokens) and .decideLong (256 tokens): GLiNER2.5-Decide (Fastino, Apache-2.0) from FluidInference/gliner2-5-decide-coreml. Up to 4 questions × 32 labels per call; fp16 packages pinned by revision and checksum.
• GLiNER2Manager.classifyConcurrently(text:heads:): several questions per call, and several calls in flight on one model.
• SortAnythingDemo: 1,000 Wikipedia abstracts sorted into categories chosen at run time (Show and Turbo modes, SORT_AUTOPLAY for recordings).
• SortDecisionsDemo: Fastino's Fast Decisions set, every question of a document answered in one call.

Numbers (Apple M5 Pro, 1,000 DBpedia abstracts)

	Core ML (fp16)	gliner2 PyTorch, Mac GPU
Time	~5.9 s	22.8 s
Peak memory	~1.0 GB	5.7 GB
Weights	0.92 GB	1.95 GB (fp32)
Accuracy	89.0%	89.0%

Identical predictions on all 1,000. Fast Decisions (public split): 62.9% average, identical to the original model on all 2,900 decisions.

Also since 0.2.0

• GLiClass decision demos and 2048 model comparison (#6); GLiClass models load from Hugging Face (#7).
• GLiNER 2.5 small / base / multilingual Core ML classifiers (#8).
• Sub-1B decision-model runtimes (#9).
• laya runtime and Tetris demo fixes (#5), package docs (#4).

Package

.package(url: "https://github.com/FluidInference/FluidUse.git", from: "0.3.0")

Pull requests: #4, #5, #6, #7, #8, #9, #13.

FluidUse 0.2.0

laya typed decisions on Core ML

• LayaManager: Swift port of laya (open Jev-style decision model, Apache-2.0). choice / score / noul answers with calibrated probabilities from one encoder pass, no generated tokens. Fixed-length Core ML buckets (128/256/512/1024 tokens × 32 options) from FluidInference/laya-coreml; fp16 or e8 (int8 embedding table, 30% smaller, same accuracy).
• LayaTokenizer: mmBERT/Gemma byte-fallback BPE, byte-exact with HuggingFace tokenizers on the conversion fixtures.
• FluidUseLaya CLI: answer, tetris, benchmark.
• LayaTetrisDemo: laya scores every candidate landing live on the Neural Engine, with a stdout console for a tmux/asitop split.
• Benchmarks.md: CUA-S1-FORMS and laya on one page, every number backed by a checked-in report.

Numbers (Apple M5 Pro)

• 99.5% of ops on the Neural Engine; 3.7 ms per short question end to end, ~15,800 Tetris decisions/min.
• Accuracy identical to PyTorch on laya's 10 published suites (3,899 questions, ≥99.8% row agreement) and on the five-task Laya-vs-Jev table; p50 5.2 ms vs 61.6 ms for PyTorch on the same CPU.

Package

.package(url: "https://github.com/FluidInference/FluidUse.git", from: "0.2.0")

Pull requests: #1, #3.

FluidUse 0.1.0

First release: local computer use on Apple silicon. Accessibility driver for native apps, Safari and Chrome; WKWebView driver; CUA-S1-FORMS context/option schema; document entity extraction; predetermined answer sheet; SwiftUI demo. Depends on FluidAudio 0.15.8 for the Core ML model wrapper.