OfflineLLM v5.1.0 — GPU Acceleration ⚡

The biggest performance release yet. Still 100% offline — zero network permissions, nothing ever leaves your device.

⚡ GPU Acceleration (Vulkan)

Run models on your phone's GPU — new opt-in toggle in Settings → Performance offloads up to the entire model to the GPU via Vulkan
Shows your detected GPU by name, with a per-layer offload slider for fine control
Automatic CPU fallback — if a GPU load fails on your device, the model loads on CPU instead of erroring out
Biggest gains on Adreno (Snapdragon) and cooperative-matrix-capable GPUs; verified full 25/25-layer offload on a Pixel 6a (Mali-G78)

Per-device CPU kernels: the app now ships 7 optimized kernel builds (armv8.0 → armv9.2) and picks the best one for your exact chip at load time — unlocking dotprod/i8mm/SVE instructions that previous versions never used
Long chats no longer slow down: conversations used to re-process the entire history every turn, grinding to 30+ second stalls by turn 10. Now each turn processes only your new message — follow-up replies start near-instantly
Prompt reading now uses all CPU cores; generation keeps to the fast cores
New Performance settings: CPU thread count, memory-mapped loading, RAM lock, experimental quantized KV cache (halves memory on long conversations)

11 Ptyxis terminal palettes — Cobalt Neon (the new default look), Nord, Synthwave, C64, Borland, and more
Catppuccin in all 4 flavors (Latte / Frappé / Macchiato / Mocha) × 14 accent colours
Dracula with 7 accent colours, plus a monochrome-accent mode
13 bundled fonts — Turret Road (new default), IBM Plex Mono/Serif, Playfair Display, Press Start 2P, and more — with an app-wide text size slider
System / Light / Dark / AMOLED with Material You accents still there if you prefer

Updated llama.cpp engine: Gemma 4 E2B/E4B models now load and run, including quantized GGUFs that previously failed with loadModel() failed
Qwen 3.5 hybrid models run with new fused kernels
Model load errors now tell you what actually went wrong (e.g. a corrupted/incomplete download) instead of a generic failure

Model imports are verified against the source file size — interrupted imports fail cleanly instead of leaving a broken model behind
Streamlined 2-step onboarding
Responses containing markdown headings are no longer cut off mid-answer
Slimmer native footprint: dead-weight libraries removed from the APK

Now requires Android 14+ (arm64-v8a, as before)
GPU acceleration is off by default — enable it in Settings → Performance and reload your model
Performance settings apply the next time a model is loaded; the first message after enabling GPU includes one-time shader compilation, then it flies
Fresh installs default to the Cobalt Neon theme + Turret Road font — everything is changeable in Settings → Appearance