BigMoeOnEdge

v0.8.1 Breaking

This release includes 1 breaking change for platform teams planning a safe upgrade.

Published 9d LLM Frameworks

View tool

✓ 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

ReleasePort's take

Light signal

editorial:auto 7d

The release updates the default dense-weight policy from Warm to anon (O_DIRECT).

Why it matters: Affects systems relying on previous dense weight loading behavior; monitor performance after upgrade.

Summary

AI summary

Default dense-weight policy changed from Warm to anon (O_DIRECT).

Changes in this release

Type	Severity	Summary	CVE
Feature	Medium	Changes default dense-weight policy to anon (O_DIRECT). Changes default dense-weight policy to anon (O_DIRECT). Source: llm_adapter@2026-07-19 Confidence: low	—
Feature	Low	Keeps "Warm" option accessible in Settings for models fitting in RAM. Keeps "Warm" option accessible in Settings for models fitting in RAM. Source: granite4.1:30b@2026-07-19-audit Confidence: low	—
Performance	Low	Improves token throughput from 0.711 to 0.998 tok/s on gpt-oss-120b with a 2000 MiB cache. Improves token throughput from 0.711 to 0.998 tok/s on gpt-oss-120b with a 2000 MiB cache. Source: granite4.1:30b@2026-07-19-audit Confidence: low	—
Refactor	Low	Defers fresh-install path to a single source of truth for the default policy. Defers fresh-install path to a single source of truth for the default policy. Source: granite4.1:30b@2026-07-19-audit Confidence: low	—

Full changelog

Flips the default dense-weight policy to anon (O_DIRECT).

On a >RAM model, anon reads the dense (non-expert) weights via O_DIRECT into private buffers, so a memory reclaim swaps them to zram (fast) instead of refaulting them from flash (slow) — and the expert cache finally has room to earn its budget. Measured 0.998 vs 0.711 tok/s on gpt-oss-120b with a 2000 MiB cache.

Warm stays one tap away in Settings for models that fit in RAM, where its page-cache prefetch is the better trade. The fresh-install path now defers to a single source of truth for the default.

Note on reproducing the published numbers: those were measured with warm. To match them, set dense weights to Warm at load in Settings.

Pre-release, like v0.8.0: the session-teardown A/B is still owed, and this default flip's cross-model A/B (Qwen, Gemma) has not been run yet. Promote with gh release edit v0.8.1 --prerelease=false once measured.

Install

BigMoeOnEdge-v0.8.1-dev.apk — dev flavor, arm64. adb install -r to keep any models already on the device.

Breaking Changes

Changed default dense-weight loading policy from Warm (page-cache prefetch) to anon (O_DIRECT), altering memory handling and performance characteristics.

View diff on GitHub

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

Share on X Share on Bluesky

Track BigMoeOnEdge

Get notified when new releases ship.

About BigMoeOnEdge

All releases →