AnythingLLM Mobile v1.2.3

What's new

• Choose your language: AnythingLLM is now available in 17 languages: English, German, Spanish, French, Hindi, Indonesian, Italian, Japanese, Korean, Dutch, Polish, Portuguese (Brazil), Russian, Turkish, Vietnamese, and Chinese (Simplified and Traditional). The app uses your phone's language by default. You can change it under Settings > App language. (#89)
• Low memory warning for on-device models: Before your first on-device chat of each session, the app compares the model's memory needs against your phone's free RAM. The estimate covers the model weights, the vision projector, the KV cache at your workspace's context window, and compute buffers. If memory is low, you get a warning so your phone doesn't close the app in the middle of a reply. While memory stays low, an amber icon shows next to the model chip. (#92)
• Clearer chat errors: Failed replies now show as an error card, and long errors are collapsed behind a "Show more" toggle. Errors stay marked as errors after you reopen a thread. Loading and error states in the model picker now sit centered at every sheet height, and the extra "Could not list models" toast is gone. (#93)

Fixes

• Fixed two on-device crashes (#90):
  • A crash when a model was unloaded while a reply was still generating. This could happen through the keep-alive timeout, a workspace change or a model switch. All native calls on a model now run in one queue, so the model is only unloaded after the reply finishes.
  • A crash when two document embeddings ran at the same time on the on-device embedder. Embeddings now run one at a time.
• Fixed a crash when tapping a reply notification for a thread or workspace that had been deleted. You now see a message instead. (#91)
• Cleaner reply notifications: Successful replies now show just the workspace and the reply text. Failed replies still say "Reply failed". (#91)

Full Changelog: 1.2.2...1.2.3