Laya-CoreML 0.1.0 release artifacts

Install the public runtime on Apple Silicon:

python -m pip install 'laya-coreml[demo]==0.1.0'

The wheel contains inference, Hub loading, the terminal Snake demo and its replay renderer. Model weights are downloaded separately. PyTorch is an optional export dependency; MLX is an optional comparison dependency. Neither is required to run the released models. See the PyPI project.

Pinned model releases

All six repositories are public. Each includes the exported Core ML package, tokenizer, configuration, model card, license, attribution and validation report. ANE packages also contain the exact host embedding and action-head tensors. They are standalone inference bundles.

For example:

hf download aac6fef/laya-multilingual-coreml-ane \
  --revision 39d6a9b3d0f67f06da74fbade6121ea134cbdb21 \
  --local-dir models/ane
laya-coreml-snake --model ./models/ane

Or load a pinned Hub snapshot directly:

import laya_coreml as laya

agent = laya.load(
    "aac6fef/laya-multilingual-coreml-ane",
    revision="39d6a9b3d0f67f06da74fbade6121ea134cbdb21",
)

The release inventory, package hashes, shapes and complete bundle sizes are in hub-release.json. The W8 bundle is about 557 MB including host embeddings and tokenizer; 129 MB describes only its Core ML body. W8 is weight palettization with FP16 compute, not integer activation quantization.

Model and package verification

Each published package was checked against its archived validated export before upload. The portable ANE bundles preserve the original Core ML body and extract six host tensors with exact element equality. Packaging-time inference then repeated the upstream golden-fixture checks:

The short exports reject fixtures beyond their input capacity. Their fitting subsets are explicit; they are not full-context equivalence claims. All use the same 0.02 probability-drift threshold. The general checkpoints cover 189/189 questions; the six bundles together cover 319 fitting comparisons, with overlap between checkpoints. See release-validation.json.

A fresh Python 3.12 environment installed the built wheel and demo extra, with no Torch, MLX or Transformers. All six local bundles loaded and produced stable predictions while Python socket connections were blocked. A public, pinned ANE Hub snapshot was then loaded, followed by a second load from cache with network connections blocked. No offline section attempted a connection. This check also exercised the Core ML symlink materialization fix described in USAGE.md. Results: wheel-runtime-smoke.json.

Version 0.1.0 was published to PyPI on 2026-09-20. Both public distribution files were downloaded and their SHA256 digests matched the local release artifacts. A second new environment then installed laya-coreml[demo]==0.1.0 directly from the public PyPI index with caching disabled. The README’s direct Hub-loading example ran successfully, returned zero output tokens, and produced the same answer on three repeat calls with network connections blocked during prediction. Both installed CLI entry points also started successfully. See publication receipt and public-install inference check.

Reproduce the release checks

The local release gate passed 64 tests, Ruff lint/format checks, and strict Twine package-metadata validation. Tests include conversion semantics, bundle cache integrity, game invariants, offline model resolution and exact replay of every board/action in the published 855-move recording.

The build backend is pinned to Hatchling 1.31.0. Its Core Metadata 2.4 output passes the release’s Twine 6 checker; an initial build using the newer default 2.5 metadata was rejected. See Hatchling’s version history for the default change. The published artifacts use the compatible build and are checked again before upload. The exact wheel and source-distribution SHA256 digests are recorded in distribution-files.json.

From a source checkout with the original validated artifacts available:

uv sync --extra convert --extra dev --extra compare --extra research --extra demo --extra publish
.venv/bin/python scripts/prepare_hub.py
.venv/bin/python -m benchmarks.release_validate models/hub/* \
  --output benchmarks/results/release-validation.json
.venv/bin/ruff check .
.venv/bin/ruff format --check .
.venv/bin/pytest -q
.venv/bin/python -m build
.venv/bin/twine check --strict dist/*

prepare_hub.py refuses to overwrite existing bundle directories. The standalone wheel check runs with scripts/wheel_smoke.py --hub-check in a fresh environment from outside the source tree, after downloading or preparing the six bundles. See SNAKE_BENCHMARKS.md for complete game-loop measurements and LAUNCH.md for the exact recorded media sources.

0.2.0 upstream synchronization

This release ports applicable prompt validation, custom noul labels, Unicode instructions, conversation-tail truncation, option-collapse diagnostics and answer-confidence reporting through upstream 4aa6761. It also publishes the temperature clamp that was present in source 0.1.1 but absent from PyPI 0.1.0. Core ML graph signatures, weights and ANE capacity checks are unchanged. Historical benchmark results above describe 0.1.0, not a new performance run. Mixed-question latency and Intel Mac support have not been validated by this sync.

Push a v<version> tag after updating pyproject.toml, laya_coreml/__init__.py and uv.lock. The Release workflow runs the reusable CI checks, verifies all three versions against the tag, builds and validates wheel/sdist, publishes to PyPI using the repository secret PYPI_API_TOKEN, and creates a GitHub release with the same distributions. CI failure prevents publishing.