These notes describe the original paid release and its limits. The current free native app requires no KWIP account or activation.
What changed
- Build models with consolidated layer, activation, conditioning, embedding, tensor-operation and merge tools, with consistent configuration and captured tensor viewers.
- Edit Parallel features, Separable convolution and Transformer structure in collapsible canvas containers with grouped named inputs and outputs.
- Configure data mapping, batching, objectives and run settings directly in Train. Connect its trained model to Evaluate or Save model, and inspect prediction tables on the component.
- Prepare named numeric and text inputs, fit character tokenization on training data, create sequence windows and train next-token models. Continue prompts with bounded generation using the saved model and tokenizer.
- Use RMS normalization, inline rotary position settings and visible model-input boundaries to build and inspect small decoder architectures.
- Save canvas versions and inspect changes since the last save. Choose whether to retain optional run details while keeping required trained-model and replay information.
- Inspect components and global Results while a flow runs. Improve large-file reads, large-cell table searching and bounded prediction capture, with clearer resource-limit behavior.
- Refresh tool lessons, toolbar organization, family artwork and Results styling, and remove obsolete current-toolbar tools.
- Correct packaged named-input training to use the verified bundled component registry throughout nested preparation validation.
- Export recorded generation replay packages using an explicit destination-folder permission, retaining required verified model and tokenizer evidence.
- Signed delta updates from v1.2.0 build 17 reduce this update to about 16 MB. If the patch cannot be used, the updater falls back to the full installer with fresh account authorization.
Compatibility at release
- Existing paid v1 licenses and permanent offline use remain valid. This release remains for Apple Silicon Macs running macOS 14 or later.
- Existing project and recorded-run formats retain their versioned interpretation. Retired data operations upgrade only where exact behavior is preserved; unsupported cases require explicit repair rather than silently changing results.
- Saved trained-model packages require their supported input/preparation contract. Text generation requires a compatible saved tokenizer and sequence model; this does not include pretrained Llama weights.
Known issues and limits
- Release acceptance uses one Mac. Additional hardware and clean-Mac coverage remain unavailable.
- The text path supports bounded numeric/token inputs and small character-level decoder workflows. Image/audio tensor preparation, heatmaps and pretrained-model downloads are outside this release.
- Memory, time, scratch-space and artifact byte limits still apply. Genuine required checkpoint/replay failures are not silently discarded.
- Delta updates are offered from supported build 17; other versions use the full installer. A full installer remains available for recovery.
- Immediately starting a run before capability discovery finishes can temporarily report an unavailable objective; retry after discovery completes.
- A table first selected during active execution may need to be reselected after the run finishes to show its captured output.
Historical build reference
Source: 27085d80e0663f7487133bcb80b10a17a752455e
Installer SHA-256: 3f16ac6ab4532e48e4d3f408c48c2c1d82b63af662b1bef969d1f13743e56c0f