Switchify Prediction v0.2.1

v0.2.1

Adds asynchronous whole-word generation to the offline SmolLM2 companion while
preserving the statistical and reranking APIs and SQLite formats. Generation
uses protocol 2, so the matching v0.2.1 workers are required. Model and tokenizer
bytes are unchanged.

The generation request carries context, prefix, session identity, up to three
excluded instant words and a result limit of up to three. Search uses eight
beams, up to eight tokens per word and 64 forward evaluations. Results are
normalized whole words with a probable following boundary. Words need not
belong to the statistical vocabulary. Empty results are valid.

The inference/reset reply deadline increases from 500 ms in v0.2.0 to two
seconds. Startup remains bounded to 30 seconds. Failure requires explicit retry. No network access is needed at runtime.

Generation is not automatically production-qualified. Corpus comparisons are
regression tests with unknown pretraining overlap. Existing reranking quality
regressions remain documented in the historical qualification reports.
Prepared CI bundles are unsigned; platform applications sign their embedded
workers through their own packaging workflows.

This companion release does not publish the Switchify PC desktop RC.

The release workflow validates all platform archives, checksums, database smoke tests and 1,000 warmed generation queries on Windows and macOS. Executables are unsigned for embedding in platform applications. Model weights remain separately checksum-pinned upstream inputs.

Switchify Prediction v0.2.0

Switchify Prediction v0.2.0

Adds an optional offline SmolLM2 companion library and CLI. It returns statistical suggestions immediately and can refine them asynchronously in an isolated worker. Existing predictor APIs, personal learning, database formats and the default statistical model are unchanged.

The neural companion is experimental and not production-qualified. Optimized Windows refinement measured 122 ms p95, but two development quality cells regressed and portable performance missed the target. Linux and macOS packages passed build and automated tests; actual model performance on those platforms remains unmeasured. See the repository's docs/smol-production-results.md for measurements and limitations.

Downloads

• switchify-prediction-* packages contain the statistical CLI.
• switchify-prediction-neural-* packages contain the companion CLI and workers for the named platform. These are separate from the statistical CLI packages.
• The model bundle contains the existing English statistical database and attribution notices.
• Each ZIP has a SHA-256 file. Verify it and run the included verify_bundle.py after extraction.

Neural model weights are not included. Follow neural/README.md to assemble the pinned local model bundle. The companion never downloads weights at runtime. Packages include license notices and explicit qualification status. Executables are unsigned, and this release does not integrate the companion into Switchify PC or publish crates to crates.io.

v0.1.0

Offline English SQLite word prediction for embedded applications and CLI use.

• Model en-aac-oanc-v1: 21,674 words and 539,151 n-grams, including the evaluated AAC/OANC corpus mixture.
• Contextual word completion, separate local personal learning, transactional imports/reset, and explicit production model validation.
• Native macOS Apple Silicon, Windows x64 and Linux x64 CLI ZIPs; platform-independent model ZIP with provenance, evaluation reports and corpus attribution.
• All release CI checks and downloaded-package smoke tests pass on macOS, Windows and Linux. Release assets and their internal checksums were verified before publication.

Download the English model ZIP and the CLI ZIP for your platform into separate directories. Check each outer .sha256, then run the included verify_bundle.py against the extracted files. Applications can consume the Rust library directly at this release commit.

CLI executables are unsigned. Linux binaries require compatible GNU libc. Personal databases are unencrypted and must stay local. Corpus terms are separate from MIT code licensing; retain the included attribution and notices, including the documented distinction between OANC's current publisher grant and historical notice. Scores and AAC test results are not clinical suitability claims.

See README and docs/production.md for usage, limitations and integration guidance.