Skip to content
Kotoshu Kotoshu 言修
News

2026-09-10 release. 10 September 2026

Language packs and the typo-biencoder hybrid in registry v1.6.0

One fetch now loads a whole language through KPK1 packs for English, German, and Portuguese, and the typo-biencoder hybrid becomes the first alternative model to beat the full tier on every gate.

  1. A language pack concatenates the dictionary files, the mini tier, the vocabulary, and the bucket table into one KPK1 artifact with per-section checksum footers, so three round trips become one. loadPack on the wasm surface and pack mode in @kotoshu/worker both fall back to per-artifact loading when no pack exists.
  2. The model bake-off closed with a winner: the 0.48 MB typo-biencoder hybrid, which retrieves with the tiny bi-encoder and rescores with fastText, beats the full tier on every real-pair component at 3 to 4 ms per lookup. English gains +6.3 points on top-5 with a confidence interval of +4.9 to +7.7 over 2,509 pairs, German gains +28.6, Spanish gains +10, and no component regresses. The hybrid is registered opt-in as kotoshu://models/typo/typo-biencoder, and the candidates rejected on measurement along the way were int4, the MiniLM rerank, ModernBERT, and the standalone bi-encoder.
  3. The registry now carries 383 release assets, and the lid primary URLs work again after pointing at unattached assets since v1.5.0.

see also models-fasttext-onnx Performance docs