KWIPKwipily

Historical release · 2026-10-22 · Build 31

Kwipily 1.3.3

Build with models, media and self-play

Prepare images and audio, classify documents, transcribe speech, train editable models and run bounded game-learning workflows on one canvas.

These notes describe the original paid release and its limits. The current free native app requires no KWIP account or activation.

What changed

  • Prepare image and PCM WAV media with bounded captured previews; classify documents and audio using supported model adapters.
  • Process images and run OCR using separately installed qualified model packs; transcribe supported audio to standard transcript and segment tables.
  • Author bounded declarative turn-based game rules, bind models to game sessions, and train policy/value models with retained replay, campaigns and evaluation.
  • Expand supported causal-model fine-tuning with prompt/response loss, parameter selection, reviewed resume and save purposes; build editable mixture-of-experts models in one canvas container.
  • Use Prebuilt models for supported classification, regression and clustering workflows, with standard table outputs and retained model contracts.
  • Load the pinned supported Laya adapter and use Decide with question sets and standard prediction tables.
  • Updated lessons, bounded table-cell inspection, retained project/History compatibility and responsiveness protections.
  • Signed updates from build 26 use an approximately 65.4 MB patch; automatic full-installer fallback remains available.

Compatibility at release

  • Apple Silicon Macs running macOS 14 or later.
  • Existing paid v1 permanent offline rights and two-Mac license scope remain unchanged.
  • Existing projects and captured History remain supported within documented schema contracts.

Known issues and limits

  • Model and device support is adapter-specific. No arbitrary downloaded model code or universal architecture support is claimed.
  • Media and model packs are separately acquired; captures, previews, replay and training campaigns have explicit bounds.
  • Declarative game rules and supported policy/value models do not provide arbitrary executable rule scripts or a general interactive lab.
  • Schedules require the app open and Mac awake; inbound webhooks remain authenticated loopback only.
  • No new color themes or TTS are included. Optional backend prerequisites and one-Mac release acceptance remain separately reported.
  • The Kmeans save receipt currently uses classifier wording; the exported clustering model and reload behavior are unaffected.
Historical build reference

Source: c7416ebb405f74cc676c25be2177910e5f5350a8

Installer SHA-256: de7dea5da23b2db19576b5fec123254827061faa72ca1e577b8f5bee83a1a166