Skip to content

Your Own AI v0.4.0

Choose a tag to compare

View all tags
@github-actions github-actions released this 15 Aug 05:26
· 31 commits to main since this release
v0.4.0
d7d4e83

Qwen 3.8-27B, offline with vision

  • Alibaba's newest open model joins the offline catalog - frontier-class coding, math, and reasoning with built-in thinking, Apache-2.0 licensed
  • Hybrid attention keeps long documents fast, with a 262K context window
  • An optional vision add-on lets it see images you attach
  • Best on 24GB-class graphics cards, or 32GB RAM on the processor

Your records now work on macOS

  • Fixed: on Mac, the record-keeping engine crashed the first time it set up an AI's conversation records - every macOS install was silently affected, and new chats were never saved to your records
  • The cause was missing security entitlements on our bundled engine; it now ships with the same set Holochain's own desktop apps use, and records set up in seconds on first launch (confirmed on Apple Silicon)
  • If your Mac ended up in "Running on CPU for stability" after force-quitting the broken app, one click on "Try GPU again" restores full speed - and macOS now tolerates force-quits without tripping that safety net
  • Huge thanks to the tester whose codesign-level bug report led us straight to it

Switch models where you are

  • The model name on every AI card is now a switcher - click it to change that AI's model in place
  • A quiet chip beside the chat's Ask row shows the current model arrangement and switches it in one tap (Settings > Appearance can hide it for a bare chat)
  • Same choices everywhere: automatic routing modes, your offline models with their fit dots, online models, or your own connected server

Honest on every graphics card

  • When a graphics driver can't actually run models, the app now detects it, switches to your processor automatically, and says so plainly - with a driver-update hint when one would genuinely help. "Model too large" is no longer misused for driver failures
  • The optional NVIDIA CUDA engine is only offered on cards it supports
  • Model recommendations, fit badges, and context sizing all plan for the processor when the graphics card is out of play
  • Windows engine downloads are now code-signed

Under the hood

  • Inference engine updated to llama.cpp b10435 - fixes Muse Glimmer occasionally losing a trailing tool call in agent work, plus sharper tool-call parsing for Qwen models
  • Automatic model picks no longer grab a bigger model while your current one is still loading - fit is judged as if the slot were free, so balanced routing stays balanced
  • On slow or busy machines, record setup now waits out the storage engine's warmup instead of showing "records couldn't be set up" too early
  • Diagnostic reports on macOS now include the actual crash cause from system crash records

Full download options and checksums below. Everything local stays free, forever.

Assets 9
Loading