Download Latest Version v.1.3.6 source code.zip (28.1 MB)
Home / 1.3.6
Name Modified Size InfoDownloads / Week
Parent folder
README.md 2026-11-12 1.7 kB
0
v.1.3.6 source code.tar.gz 2026-11-12 27.7 MB
0
v.1.3.6 source code.zip 2026-11-12 28.1 MB
0
Totals: 3 Items   55.8 MB 0

New on-device engine

Android 182, iOS 42.

1.3.6 is 1.3.5 moved off the /legacy runner of react-native-executorch onto the RNE 0.11 runner. Nothing else from main is in it; the redesign ships in 1.4.0.

  • Chat models run on the RNE 0.11 runner instead of /legacy, and the react-native-executorch legacy patch is gone
  • Models downloaded under the old engine are marked as needing an update and re-download the v0.11 exports: an Update button on the model card, a sheet on each new chat, a notice at the top of an affected chat, and a clear message when offline. After an update the new files are loaded without restarting the app
  • LLaMA 3.2 QLoRA is retired; its chats move to LLaMA 3.2 SpinQuant
  • Documents are embedded with the RNE 0.11 text embedder, and dictation transcribes with the RNE 0.11 Whisper pipeline
  • The first dictation pauses other model downloads while Whisper is fetched (#490)
  • Whisper's non-speech notes ([BLANK_AUDIO], (mumbling)) are dropped from the dictated text
  • The <answer> wrapper LFM 3 VL puts around its reply is stripped
  • Gemma 4 VL on Android loads from the v1.0.0 Vulkan export
  • Qwen 3 is no longer offered on iPhone SE
  • Fixes: the keyboard stays down under the search- and document-model download sheets, and a photo earlier in a chat no longer breaks a text-only model

Known issues: dictation occasionally fails to start with "Could not start live transcript". RNE 0.11 Whisper decodes greedily, without the confidence fallback the legacy runner had; that lands in 1.4.0.

Release branch: release/1.3.6-store; recorded on main in [#531].

Full Changelog: https://github.com/software-mansion-labs/private-mind/compare/1.3.5...1.3.6

Source: README.md, updated 2026-11-12