Erebrus AI 1.1.2

Version: 1.1.2+4

Artifact	Platform
erebrus-ai-dappstore-v1.1.2.apk	Solana Mobile / dApp Store (dApp Store key)
erebrus-ai-linux-v1.1.2.tar.gz	Linux bundle (x64)
erebrus-ai-playstore-v1.1.2.aab	Android App Bundle for the Play Console (Play Store key)
erebrus-ai-playstore-v1.1.2.apk	Android sideload (Play Store key)
erebrus-ai-r8-mapping-v1.1.2.zip	R8 mapping files for deobfuscating Android stack traces
erebrus-ai-windows-v1.1.2.zip	Windows
erebrus-ai-manifest-v1.1.2.json	Release manifest (sizes, SHA-256, signing certificates)
erebrus-ai-sha256sums-v1.1.2.txt	Checksums (sha256sum -c erebrus-ai-sha256sums-v1.1.2.txt)

Both APKs use com.erebrus.ai with different certificates — install one.

Changes

Model catalog

• Supports the 60-model catalog (2026.11.02). Onboarding and the model
  browser now follow the catalog's per-device recommendations for phones,
  laptops, desktops and gaming rigs, falling back to smaller bands when a
  pick does not fit the device's memory.
• Recommendations only name models the device can actually run. Previously
  a 32 GB Windows or Linux PC could be recommended an experimental model it
  could not download.
• Phones no longer offer a download for models that cannot fit in memory;
  the card shows TOO LARGE and the memory the smallest package needs.
• Recommendations list official/stable models first, then experimental
  packages, then uncensored (reduced-safety) community builds. Uncensored
  builds are labelled UNCENSORED and ask for confirmation before
  download in onboarding and the model browser. Experimental packages are
  labelled.
• Models whose only package needs a runtime this build does not include
  (such as the PrismML llama.cpp fork) are hidden instead of failing after
  download.
• Importing a GGUF that uses a quantization type the bundled runtimes cannot
  load (for example Q2_0) is refused up front. On Windows and Linux the
  TurboQuant runtime would otherwise misread such a file.
• Desktop: the download row no longer overflows for models that support GPU
  offload.

Voice

• Two new local speech engines in Settings → Voice, running on sherpa-onnx:
  Parakeet (NVIDIA Parakeet TDT 0.6B v3, about 670 MB, 25 European
  languages with punctuation) and Moonshine (Tiny or Base, about
  120–290 MB, English only, sized for phones). Each download is pinned to an
  immutable revision and checked byte-for-byte against its SHA-256 before
  use.
• Automatic keeps the existing default — on-device system speech first —
  and, where that is unavailable, uses the best downloaded model for the
  device and language: Parakeet on laptops, desktops, gaming rigs and 12 GB+
  phones, then Whisper (Moonshine for English on phones).
• Choosing Parakeet, Moonshine, Whisper or System speech explicitly always
  wins when it can run. When it cannot (not downloaded, wrong language), the
  recording falls back to Automatic and the Transcribe screen says why,
  instead of failing.
• Parakeet and Moonshine recordings are split by voice activity, so saved
  transcripts carry per-utterance timecodes. Decoding runs off the UI thread
  and Cancel stops it immediately.
• The bundled sherpa-onnx and ONNX Runtime libraries add about 25 MB on
  iOS, 27 MB per Android ABI and 55 MB to the universal macOS app.