Skip to content

BigMoeOnEdge

v0.9.0 Feature

This release adds 3 notable features for engineering teams evaluating rollout.

Published 8d LLM Frameworks
✓ No known CVEs patched
Read the diff → Tool health → What is this tool? →

✓ No known CVEs patched in this version

Topics

android c++ edge-ai gemma gguf gpt-oss
+7 more
inference llama-cpp llm mixture-of-experts moe on-device-ai qwen

Summary

AI summary

In‑app model downloads now use internal storage with O_DIRECT, enabling direct I/O and resumable transfers.

Changes in this release

Feature Medium

In-app downloads now use O_DIRECT-capable internal storage, enabling direct I/O for large models.

In-app downloads now use O_DIRECT-capable internal storage, enabling direct I/O for large models.

Source: llm_adapter@2026-07-19

Confidence: high

Feature Medium

Download resume supports HTTP Range requests, continuing from last offset without restarting transfer.

Download resume supports HTTP Range requests, continuing from last offset without restarting transfer.

Source: llm_adapter@2026-07-19

Confidence: high

Feature Medium

Downloads no longer create a temporary second copy, requiring only model‑size free space.

Downloads no longer create a temporary second copy, requiring only model‑size free space.

Source: llm_adapter@2026-07-19

Confidence: high

Feature Medium

Rename-on-complete ensures half‑finished downloads are not listed as runnable.

Rename-on-complete ensures half‑finished downloads are not listed as runnable.

Source: llm_adapter@2026-07-19

Confidence: high

Bugfix Medium

UI download progress shows MiB counter on its own line, preventing ellipsis of long model names.

UI download progress shows MiB counter on its own line, preventing ellipsis of long model names.

Source: llm_adapter@2026-07-19

Confidence: high

Bugfix Medium

Cancel button no longer races with ongoing download, avoiding inconsistent states.

Cancel button no longer races with ongoing download, avoiding inconsistent states.

Source: llm_adapter@2026-07-19

Confidence: high

Bugfix Medium

Download progress bar clears when no download is in flight.

Download progress bar clears when no download is in flight.

Source: llm_adapter@2026-07-19

Confidence: high

Bugfix Medium

Fixes issue #67 where in‑app downloads lost streaming advantage due to buffered I/O fallback.

Fixes issue #67 where in‑app downloads lost streaming advantage due to buffered I/O fallback.

Source: llm_adapter@2026-07-19

Confidence: high

Bugfix Low

Known limitation: Android 14 may block background WorkManager restarts after app kill; download resumes only on app reopen.

Known limitation: Android 14 may block background WorkManager restarts after app kill; download resumes only on app reopen.

Source: llm_adapter@2026-07-19

Confidence: low

Full changelog

In-app downloads now use O_DIRECT-capable internal storage

Downloading a model in the app — from the catalog or a pasted URL — used to write to the app's external files dir, an emulated/FUSE volume where O_DIRECT silently returns wrong data. The engine caught it and fell back to buffered I/O, so every in-app-downloaded model lost the streaming advantage the engine exists for (measured on device: o_direct=0). The system DownloadManager cannot write to internal storage by design, so it is replaced.

DownloadWorker is a foreground WorkManager transfer that streams the gguf over HTTP straight into filesDir/models (real f2fs, where O_DIRECT works — the same dir the file picker imports to). Validated end-to-end on device (OnePlus 15R): a ~20 GB model downloaded in-app now loads in direct (O_DIRECT) mode.

What changed

  • Downloads land on internal storage → real O_DIRECT, not buffered. Fixes #67.
  • Resume: an interrupted .part resumes from its offset with a Range request, following Hugging Face's resolve → CDN redirect manually so the header survives (instead of restarting a multi-GB transfer from zero).
  • No temporary second copy: a download needs free space equal to the model size, not double.
  • Rename-on-complete, so a half-finished download is never listed as runnable.
  • UI: the MiB counter sits on its own line so a long model name can't ellipsize it away; the bar clears when nothing is in flight; Cancel no longer races back to "downloading".

This mirrors the mechanism in Google's own AI Edge Gallery (HttpURLConnection + Range + WorkManager foreground worker).

Known limitation (#69): if the OS kills the app while it is in the background mid-download, Android 14 blocks WorkManager from restarting the foreground transfer until you reopen the app. The .part is preserved and resumes when you re-trigger the download — no bytes are lost, but it does not continue on its own.

Install: sideload app-dev-debug.apk (dev flavor; keeps all-files access so it can also read adb-pushed models). Models already downloaded by older builds keep working.

Weekly OSS security release digest.

The CVE patches and breaking changes that affected production tools this week. One email, every Sunday.

No spam, unsubscribe anytime.

Share this release

Track BigMoeOnEdge

Get notified when new releases ship.

Sign up free

About BigMoeOnEdge

All releases →

Beta — feedback welcome: [email protected]