Skip to content

BigMoeOnEdge

v0.11.1 Maintenance

This release keeps dependencies and maintenance posture current for teams operating this tool.

Published 7d 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

Minor fixes and improvements.

Changes in this release

Deprecation Low

Corrected stale documentation references to DownloadManager and renamed field.

Corrected stale documentation references to DownloadManager and renamed field.

Source: llm_adapter@2026-07-19

Confidence: high

Refactor Low

Moved per-token metrics block from Session::generate() into GenTally.

Moved per-token metrics block from Session::generate() into GenTally.

Source: llm_adapter@2026-07-19

Confidence: high

Refactor Low

Moved telemetry breakdown logic into pure function breakdown().

Moved telemetry breakdown logic into pure function breakdown().

Source: llm_adapter@2026-07-19

Confidence: high

Refactor Low

Changed Android model download tracking to a stream via WorkManager, removing 700 ms polling loop.

Changed Android model download tracking to a stream via WorkManager, removing 700 ms polling loop.

Source: llm_adapter@2026-07-19

Confidence: low

Full changelog

A structural pass over everything merged since the repo went public. No user-facing behaviour changes — the byte-identity gates pass unchanged, and the engine produces the same numbers it did in 0.11.0.

Engine

The per-token metrics block moved out of the middle of Session::generate() into a GenTally that owns the streaming cursors and the run totals. generate() loses ten measurement locals and about fifty lines of inline bookkeeping; what a token costs is now written down in one place, while what it says stays with the generation loop.

Android example

  • Download tracking is a stream, not a poll. ModelDownloader.events() exposes progress and outcomes over WorkManager's own flow, finalizing a landed .part before reporting it. The add-model card used to own a 700 ms polling loop, do its own finalization, and re-seed its state by hand after every enqueue and cancel — all of that is gone, and a download started before the app was killed is picked back up on the first emission.
  • The telemetry breakdown is a pure function. The live-vs-average choice, the wall-additive residual, and the CPU-occupancy arithmetic moved out of the panel into breakdown(), next to the contract it implements.

Documentation

Stale references corrected: the model catalog and the add-model section still named the system DownloadManager, retired when in-app downloads moved to WorkManager, and one KDoc link pointed at a field that had been renamed. RunInfo::dense_weights no longer advertises a "warm" default the engine stopped defaulting to.

Install

app-dev-release.apk below is signed with the same key as every release since 0.9.1, so it installs over an existing build without uninstalling — your downloaded models stay put.

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]