Repository navigation
Releases
Releases · LucknowAI/loki_llm
v1.2.0 — Multimodal: Gemma 4 on-device vision
On-device image-grounded chat: attach an image to a chat message and get a text response grounded in it, from a vision-capable model running fully on-device (no cloud). See CHANGELOG.md for the full list.
Added
- Render chat messages as Markdown (#39, thanks @ItsPriyamSri)
- Agent tool:
get_settings(#12) - Multimodal foundation: mmproj vision projector + attached images,
mtmdwired intollama_engine(#19, #20) - Multimodal chat UI: image attachment picker, preview, cleanup (#21)
- Gemma 4 E2B & E4B catalog entries + prompt templates (#22)
- Vision-grounded generation for plain chat and agent mode (#24)
- RAM guard + friendlier OOM messaging for vision models (#25)
- Models screen loading overlay + interaction lockout (#61)
Fixed
- Agent Stop correctness for tool-enabled turns (#6, thanks @Zedrider9t)
- Android
.ggufsideload crash;.tasksideload rejected (#7, #40) - Native SIGABRT after agent tool call (#43)
- Gemma 4 GGUF load failure — llama.cpp bumped to b10775 (#52)
- Models screen loading-state and concurrent-load bugs (#61)
- Chat composer clear/retry-wedge bugs (#59)
- Retry-load overlay wedge and mmproj HTTP 416 retry loop (#64)
- Tool settings visibility for pre-1.2.0 users (#67)
send()timeout no longer races a retry (#68)
Contributors
Thanks to our external contributors for this release: @ItsPriyamSri (#78, #79) and @Zedrider9t (#48)
Full Changelog: v1.1.0...v1.2.0
v1.1.0
v1.1.0
Second stable release of Loki LLM — on-device AI chat for Android.
Install
- Download
loki-llm-v1.1.0-android-arm64.apk(recommended for most phones). - Optionally use
loki-llm-v1.1.0-android-arm32.apkfor older 32-bit devices. - Enable Install unknown apps for your browser/files app if prompted.
- Open the APK and install.
- In the app, download a model from the catalog — models are not bundled in the APK.
Downloads
| File | Device |
|---|---|
loki-llm-v1.1.0-android-arm64.apk |
Most modern Android phones |
loki-llm-v1.1.0-android-arm32.apk |
Older 32-bit ARM devices |
Added
- Agent tools:
recall_memory,list_documents,unit_convert - Lazy GBNF grammar for more reliable tool-call JSON
- Per-tool enable toggles in Settings
- Unified RAG + agent mode in chat
- Per-iteration fields in Model I/O trace
- Agent mode gating for Gemma 270M
- Export/share conversations as Markdown
- Per-message copy on chat bubbles
- Read-aloud on assistant messages (Android)
SpeechTextNormalizerfor TTS- Model load feedback (SnackBar, active row, chat banners)
- OSS docs: CONTRIBUTING, CODE_OF_CONDUCT, SECURITY, issue/PR templates
Fixed
- Agent-mode native crash during multi-step generation
- Model load SnackBar showing wrong model on switch
- Model switch now unloads prior backend correctly
Full changelog: https://github.com/LucknowAI/loxi_llm/blob/v1.1.0/CHANGELOG.md
v1.0.0
Changelog
All notable changes to Loki LLM are documented here.
Format based on Keep a Changelog.
[1.0.0] — 2026-08-01
First stable on-device release.
Added
- On-device chat with true incremental streaming and Stop button (Android)
- Per-model chat templates (Gemma, Phi, Llama)
- Model hub: curated GGUF catalog, Hugging Face downloads, sideload
- RAG: PDF / DOCX / TXT ingest, BGE-small embeddings, ObjectBox HNSW
- Tool-calling agent mode with
calculator,datetime,document_search - Rolling conversation memory (summary beyond recent window)
- Model I/O trace diagnostics (optional JSONL log)
- In-app log viewer
llama_engineplugin (MIT, vendored llama.cpp)
Fixed
- Flutter analyze: const constructors in
ChatStatedomain tests