Skip to content

Blobs 0.5.0

Latest
Latest

Choose a tag to compare

View all tags
@nonomnonom nonomnonom released this 26 Oct 15:54
· 16 commits to main since this release
v0.5.0
843a209
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Blobs 0.5.0 — Production Release

Fast-build turn wins, sources on screen, Qwen2.5 0.5B in the list.

Performance

  • Median turn time: 97.6s → 25.5-28.9s on 8GB emulator (3 independent captures)
  • Prompt bound: real 6,000-character limit with history shedding before passage cuts
  • Answer bound: 320 → 200 tokens for all modes (product owner decision 2026-10-25)
  • Context capacity: fixed to prevent KV cache doubling per turn
  • Turn attribution: every stage timed in native code (model load, template, tokenize, context prepare, prefill, grammar, sampler, first token)

Features

  • Chat now shows retrieved sources on screen while the model writes
  • Download list revision 2026102502 adds Qwen2.5 0.5B Q4_K_M for 2GB-class devices
  • Per-stage native timings carried across JNI and written into device captures
  • Prompt sheds retained history before cutting passages, so squeezed shelf still answers

Installation

  1. Download app-release.apk below
  2. Open on Android 9+ device (ARM64 or x86_64)
  3. Allow installation from browser/file manager when prompted
  4. Open Blobs, download recommended model, then download starter library
  5. Ask your first question — works fully offline after setup

Upgrading from 0.4.0: installs in place, no data loss
Upgrading from 0.3.14 or older: export conversations first, uninstall, then install 0.5.0 (signing key changed in 0.4.0)

Verification

APK SHA-256: e3e1ab7a9677357405d8ab059b5dc1daff776f8d82dc6f0beb09856aec604b06

Certificate SHA-256: 77:81:8F:69:02:8E:7D:5E:7E:D1:C3:79:1E:1C:CF:91:63:62:06:9A:6A:3F:37:39:31:5F:80:B5:84:E3:F6:14

All CI checks passed: Android build, corpus tests (84 tests), documentation build, formatting standards.

What's New Since 0.4.0

68 commits including:

  • Speed governor wave-a: per-stage native timings, fixed context capacity, bounded prompt and answer
  • Offline research engine: fast-build, concurrent retrieval, sources on screen
  • Download list revision 2026102502: Qwen2.5 0.5B Q4_K_M for 2GB-class devices
  • Prompt bound increased to real 6,000 characters with history shedding
  • Answer bound reduced 320 → 200 tokens for all modes
  • Chat shows retrieved sources while model writes

Full Changelog

See CHANGELOG.md for detailed technical notes.

Assets 4
Loading