On-device AILive
Edge Lab
Open on-device experiment matrix for Gemma / LiteRT-LM on iPhone. Four presets, versioned JSON manifests, no cloud.
4 presets · JSON manifest · BYOM
Started Jun 2, 2026
A lab instrument for Gemma on iPhone — not a Gallery clone. Import your .litertlm, run four presets, export a versioned JSON manifest with decode tok/s, TTFT, requested vs actual backend, and thermal state. Fully local. No cloud. No hidden toggles.
Reference run · iPhone 16 Pro Max
gemma-4-E2B-it.litertlm · LiteRT-LM 0.12.0 · iOS 26.6 · 256 decode tokens per preset
| Preset | Decode Speed | Wall Time | Backend |
|---|---|---|---|
| Greedy GPU | 14.6 tok/s | 10s | gpu |
| Sampled GPU | 39.8 tok/s | 7s | gpu |
| Greedy CPU | 4.7 tok/s | 52s | cpu |
| Sampled CPU | 4.9 tok/s | 60s | cpu |
Why this exists
The Gemma 4 Benchmark Suite answers "how smart is this model on desktop hardware?" Edge Lab answers "how fast is this exact .litertlm on your phone, with settings you can reproduce?" Same model family, different instrument — quality scoring vs edge throughput manifests.
Features
- Four named presets: Greedy vs sampled × GPU vs CPU. Names match what's logged — not opaque Gallery UI labels.
- Backend honesty: requested_backend vs backend plus did_fallback when artisan-only weights force GPU on a CPU run.
- Share sheet export: Copy X thread text, JSON, or Markdown from the device. Drop manifests into GitHub or a blog.
- BYOM: Any compatible .litertlm from HuggingFace litert-community or your own builds.
- vs AI Edge Gallery: Google's demo is closed iOS. Edge Lab is Apache-2.0 with every knob in the export.
- Not a benchmark suite: No LLM judge, no 403-test scoring grid — one matrix run, one manifest, ship the numbers.
iOSLiteRT-LMGemma 4Local-first