Download Latest Version v.1.3.4 source code.zip (28.3 MB) Google Add to Preferred Sources
Home / 1.3.1
Name Modified Size InfoDownloads / Week
Parent folder
README.md 2026-10-09 8.6 kB
0
v.1.3.1 source code.tar.gz 2026-10-09 27.9 MB
0
v.1.3.1 source code.zip 2026-10-09 28.3 MB
0
Totals: 3 Items   56.2 MB 0

Highlights

  • The model in the header is the model that answers. Picking a model while another was still loading, or tapping a suggestion mid-switch, could let the previous model answer. A turn now waits for the switch and runs on the model the chat shows (#415).
  • Leave mid-answer and the reply is still there. A turn keeps running while the app is in the background, and stopping one always says it stopped instead of going quiet (#427, [#373], [#370]).
  • Model downloads carry on in the background. react-native-executorch 0.9.2 → 0.10.4 (legacy API) brings resumable background downloads; only the current download owns the progress state (#389, [#409]).
  • Documents and web search side by side. Web search is held off only for the message a document is sent with, not for the rest of the chat (#426).
  • Faster generation on iOS. The measured causes of generation slowing to a crawl on iOS are gone (#425).

Fixes

Chat

  • A repetition loop is cut at the repeat, not at the text it repeats (#374).
  • The composer button morphs between mic, send and stop, and shows busy while a model loads or switches (#408, [#395]).
  • A new chat arrives in one movement; the scroll-to-latest button has an entrance, exit and press (#404, [#403]).
  • Generated titles are cut between words; the chat title gets its full width (#400, [#385]).

  • Suggestion cards end where their prompt ends (#399); the background gradient survives a screen remount (#390).

  • The empty chat's What's new card stays below the header: its badge and title stay put and only the list scrolls, with the scrollbar in the card's margin.

Dictation

  • Dictation stops when the app leaves and keeps what was heard, releases the audio session and says why it ended, and sends the transcript the sheet shows.

Models

  • Only starting models the device can run are offered, and a family with no runnable variant says so at the list level (#421, [#396]).

Attachments and onboarding

  • The image attachment flow was reworked into an in-app picker (#310).
  • Onboarding is one surface from intro to last slide (#337).

Under the hood

  • Streaming no longer re-renders the composer, the closed drawer or every model card per token (#353, [#346], [#348], [#347], [#350]).
  • The benchmark is reproducible and reports real peak memory, including on Android (#349, [#410]); token rates are reported only when measured (#398).
  • The Android release workflow installs platform-tools explicitly and retries yarn install.
  • The data-sync foreground service that react-native-background-downloader merges into the manifest is removed: models download through expo-file-system, so the service was never started.

What's Changed

Full Changelog: https://github.com/software-mansion-labs/private-mind/compare/1.3.0...1.3.1

Source: README.md, updated 2026-10-09