KWIPKwipily

Historical release · 2026-10-20 · Build 26

Kwipily 1.3.2

Fine-tune models and connect your workflow

Load supported models, prepare examples, inspect training and save the result on a visual canvas, with bounded tensor heatmaps and connected data workflows.

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

What changed

  • Training charts report within epochs, with metric inspection and consistent live/History presentation.
  • Inspect captured tensors as Values or bounded Heatmaps, including retained post-run playback.
  • Load supported local or Hugging Face causal models with matched tokenizers; fine-tune supported models and save complete models or adapters.
  • Write tables to supported databases with column mapping and write receipts; read and deliver tables through configured APIs.
  • Configure app-open scheduled runs and authenticated local webhooks in project and global settings, with saved-version execution and delivery history.
  • Updated lessons, settings/navigation, data pipelines and retained project compatibility; improved generation admission and added canvas scroll zoom sensitivity.
  • Correct packaged database/API access, default merged LoRA export and automatic generation handoff while inspecting results.
  • Signed updates from build 21 use an approximately 22.6 MB patch; automatic full-installer fallback remains available.

Compatibility at release

  • Apple Silicon Macs running macOS14 or later.
  • Existing paid v1 licenses, permanent offline use and two-Mac rights are unchanged.
  • Saved projects, canvas versions and captured History retain their documented compatibility.

Known issues and limits

  • External models require verified supported adapters and safe weights/tokenizers; arbitrary repository code, GGUF and Laya are not supported.
  • Schedules require Kwip ML open and the Mac awake. Webhooks accept authenticated loopback requests only.
  • Heatmaps and playback use bounded captured summaries rather than retaining every full tensor.
  • One-Mac release acceptance; optional backend prerequisites remain separately reported.
  • LoRA continuation is not supported. External causal-model fine-tuning requires compatible GPT-2 or Llama architecture, safe checkpoint files and a matching tokenizer.
  • Memory and storage requirements depend on model size, training mode and workflow.
Historical build reference

Source: 48002f9ca55604412b7c2bd27d4aad8ee523ca3b

Installer SHA-256: bc3faea57edbc8b61629ee9ab210895e69ba4054f7cf939cd707b9f5a6dc4510