Historical release · October 24, 2026 · Build 34
Version 1.3.4 — Let AI build tools without code.
Kwip ML becomes Kwipily. Work with an AI assistant through MCP, inspect a visible canvas and reuse a workflow when new data arrives.
- Business starting points. Open templates from the project page, blank canvas or File menu, then connect compatible inputs and report destinations in the setup editor.
- More dependable MCP journeys. Use staged terminal actions, bounded result reads and explicit project, captured-value and History permissions. Recorded walkthroughs retain their original development-build context.
- Shared data execution. Columnar batches, segmented retained tables and fused operations reduce repeated row processing and serialization while retaining live inspection and History.
- Reusable fitted preparation. Save supported learned preparation with Prebuilt models and reuse the verified package on compatible new data. API and database outputs retain their verified external-write boundary.
Eight fresh template workflows passed release checks. Customer purchase grouping can stop at an input-integrity safeguard before producing results and remains a known issue. Other limits include asset-inventory warnings, imported-block visual status and generated Code scopes. OCR output needs review. The macOS menu bar may still use Kwip ML; naming cleanup follows in a later update.
Existing projects, licenses, two-Mac rights and permanent offline activation are preserved. Technical identifiers and the Kwip ML.app payload filename stay stable. Signed updates from build31 use an approximately11.24MB patch, with actual automatic full-installer fallback verified. Multi-million-record capacity and universal sub-five-second runs are not qualified.
Released October 22, 2026 · Build 31
Version 1.3.3 — Models, media and self-play.
Prepare more kinds of local data, train supported models and inspect their results through the shared canvas and table viewers.
- Documents and media. Classify documents, prepare images and audio, extract text with OCR, transcribe speech and classify audio using supported model packs.
- More model choices. Configure Prebuilt models for classification, regression or clustering; build editable mixture-of-experts models; and use the qualified Laya decision workflow.
- Fine-tuning and game learning. Use expanded Train controls, explicit save purposes and bounded self-play campaigns against configurable game definitions, with retained replay and independent evaluation.
- Consistent inspection. Predictions and scores use standard tables. Open long cells in a bounded, paged viewer and follow updated lessons beside the canvas.
Supported architectures, formats and devices are explicit. Model packs may require separate downloads; arbitrary repository code and universal model compatibility are not included. Game sessions are bounded experiments, not a claim of competitive playing strength.
Complete 1.3.3 release notes, compatibility and known limits. Signed updates from build 26 use an approximately 65.4 MB patch, with verified automatic full-installer fallback. The minor Kmeans save-receipt wording issue is retained for the next update; exported clustering packages and reload behavior are unaffected.
Released October 20, 2026 · Build 26
Version 1.3.2 — See training. Connect your workflow.
Inspect training as it happens, work with supported pretrained models, and connect local workflows to configured data destinations.
- Training charts. Follow loss and captured metrics within an epoch, inspect points and use the same presentation for live runs and recorded History.
- Tensor heatmaps. Keep the Values viewer or switch to bounded heatmaps, with playback of retained frames after training.
- Load and fine-tune models. Import supported safe local or Hugging Face models with matched tokenizers, train full or selected parameters or supported LoRA adapters, and save models or adapters for reuse.
- Database and API connections. Map output columns, choose supported write modes, read or deliver API tables, and inspect bounded write/delivery receipts.
- Local automation. Configure app-open schedules and authenticated loopback webhooks from project or global settings, with pinned saved versions and job/delivery history.
- Everyday improvements. Updated lessons, settings and recent-project handling, dependable imported generation after reopening, and adjustable canvas scroll zoom sensitivity.
Supported model adapters and safe file formats remain explicit. Arbitrary model repository code, GGUF, Laya, public/LAN webhooks and after-quit scheduling are not included. Heatmaps retain bounded summaries rather than every full tensor.
Complete 1.3.2 release notes, compatibility and known limits. Signed updates from build 21 use an approximately 22.6 MB patch, with automatic full-installer fallback.
Released October 18, 2026 · Build 21
Version 1.3.1 — Learn beside your canvas.
Follow lessons without leaving your workflow, with clearer reading controls and optional access for your AI client.
- Lessons beside the canvas. Guided builds, examples, practices and construction assessments use a shared lesson rail and consistent content blocks. Relevant components are marked as you move through steps.
- Easier reading and navigation. Adjust lesson text size, click colored component references to focus them on the canvas, and return to the Library from the final step.
- Keep your place. Save or unsave lesson bookmarks, see library progress markers, and start a fresh practice without losing your previous attempt.
- A clearer first Foundations practice. Learn what its example data and model represent, how batch and sample shapes differ, and which controls to change. Other lessons remain subject to individual content review.
- Training and run feedback. Training-side View tables are captured during runs. Model output checks and named-input findings appear in context, Run reflects configuration readiness, and saving gives visible feedback.
- Optional local MCP. Connect a compatible AI client to permitted projects for tool discovery, canvas editing, execution and result inspection. It starts off; project, action, component and captured-value permissions stay under your control. See optional AI client access.
Complete 1.3.1 release notes, compatibility and known limits
Released October 16, 2026 · Build 20
Version 1.3.0 — Build models. Train on text.
Build and inspect model architectures, prepare text for training, and continue a prompt with your trained model.
- Focused model tools. Consolidated layers, activations, normalization, dropout, embeddings, tensor operations and merge tools use consistent configuration and tensor viewers.
- Editable containers. Expand Parallel features, Separable convolution and Transformer containers on the canvas, move components in and out, and manage named inputs and outputs.
- Training in one place. Configure data mapping, batches, objectives and run settings in Train. Connect its trained-model output to Evaluate or Save model, and inspect predictions as tables.
- Text to tensors. Prepare named numeric and text inputs, fit character tokenization on training data, and create sequence windows for next-token training. Inspect tokens alongside their IDs and continue prompts with bounded generation.
- Decoder building blocks. RMS normalization, rotary position settings and visible Model input boundaries support small Llama2-style architectures. Pretrained Llama weights are not included.
- Canvas history and capture controls. Save canvas versions, review changes since the last save, and choose whether to retain optional run details while preserving required model and replay information.
- Smaller updates. The signed update from 1.2.0 build 17 is about 16 MB, compared with the 1.26 GB full installer. The updater falls back to the full installer when a patch cannot be used; other starting versions use the full installer.
- Workflow refinements. Inspect Results during a run, use refreshed lessons and tool organization, and benefit from improved large-file reads and clearer resource-limit handling.
Complete 1.3.0 release notes, compatibility and known limits
Released October 14, 2026 · Build 17
Version 1.2.0 — Prepare, explore and save data.
A broader local data workflow, from configurable readers to complete-table exploration and export.
- Table Input. Read CSV, Parquet, JSON/JSON Lines, Excel, Markdown, XML and HTML tables with format-specific controls, or read a specific SQLite, PostgreSQL or MySQL table.
- File Input. Read individual files or folders as tables containing metadata, document text or image pixels, with page and frame controls.
- Data preparation. Polars-backed processing powers Columns, Filter, Formula, Partition, Preprocess, Join, Union and Summarize. Sort, Unique, Reshape Table and Split Text add ordering, duplicate handling, pivots and text extraction.
- Explore tables. Choose working data for exploration, then page, search and sort complete captured tables in View. Inspect typed date/time values and image cells.
- Table Output. Export complete selected data to CSV/TSV, Parquet, JSON/NDJSON or Excel, with precision controls and optional copied-image bundles.
- Code and learning. Inspect supported preparation, model and training code, export workflow packages, and explore refreshed lessons and editable practices.
This release uses full installers and adds the client groundwork for the delta updates delivered in 1.3.0.
Complete 1.2.0 release notes, compatibility and known limits
Released October 12, 2026 · Build 14
Version 1.1.0 — A clearer canvas and focused results.
- Consistent canvas blocks. Rounded table-style components use clear headers and icon-anchored ports, with inputs on the left and outputs on the right.
- Smoother navigation. Improved interaction responsiveness and mouse-wheel zoom, plus right-button drag or Control-drag to pan.
- Cleaner light mode. White surfaces and stronger component-family borders make blocks easier to distinguish.
- Results follow your selection. Inspect a component or connection, or clear the selection for global results. Port details and connection information stay in context.
- Clearer guidance. The inspector shows the tool name above Configuration, links to its lesson in Information, and presents diagnostics with Observed, Meaning and Try next sections.
Complete 1.1.0 release notes, compatibility and known limits
Released October 11, 2026
Version 1.0.0 · build 12
The first direct Mac release.
- Build connected visual machine-learning workflows and inspect generated code, run history and results locally.
- Explore guided learning paths using the app’s sample projects and datasets.
- Use a personal v1 license on two Macs, with permanent offline use after paid activation and an exportable recovery license for the same Mac.
- Check for signed v1 updates from the app, with automatic checking controlled in Settings.
Complete 1.0.0 release notes, compatibility and known limits
Before installing
The free native app requires an Apple Silicon Mac running macOS 14 or later. No KWIP account or activation is required. Save your projects before updating and review the version-specific limits above.