litert-community/VibeThinker-1.5B
This model provides a variant of WeiboAI/VibeThinker-1.5B that is ready for deployment on Android using the LiteRT-LM.
Use the model
Android
Edge Gallery App
Download or build the app from GitHub.
Install the app from Google Play.
Follow the instructions in the app.
To build the demo app from source, please follow the instructions from the GitHub repository.
Performance (on device, measured)
Community measurement on a physical Samsung Galaxy S26 (SM-S942Q, Snapdragon 8 Elite Gen 5 / SM8850, Android 16): litert_lm_advanced_main from the litert-lm v0.16.0 release, GPU backend OpenCL (LITERT_CL) against CPU (XNNPACK), one fixed 205-token prompt text (203 tokens under this tokenizer), --benchmark. Two runs per backend taken back-to-back — cells show the range; this model thinks at length (the benchmark generations ran 3,500–3,900 tokens), so the second run starts on a warm phone and the spread is thermal, not noise. Peak RSS is the process VmHWM. Before quoting, the file was run on each backend with a real prompt and generation-checked; the GPU rows are full delegation (decode 1190/1190 and prefill 1299/1299 ops on LITERT_CL).
| Backend | Prefill (203 tok) | Decode | Time-to-first-token | Peak RSS |
|---|---|---|---|---|
| GPU (OpenCL) | 301–612 tok/s | 18.4–23.8 tok/s | 0.37–0.73 s | 735–765 MB |
| CPU (XNNPACK) | 294–310 tok/s | 18.7–21.3 tok/s | 0.71–0.74 s | 2245–2265 MB |
What the table says:
- The clearest GPU win on this bundle is memory: 735–765 against 2245–2265 MB peak RSS, 3.0× lower — the difference between fitting comfortably on an 8 GB phone and pressuring it.
- Cool, the GPU prefills ~2× faster than the CPU; warm, the two converge (the 301–612 spread is thermal). Decode is a wash either way — bandwidth-bound, the usual shape on phone GPUs.
- The long thinking traces make sustained decode the number that matters: ~19–24 tok/s means a 3.5–4k-token reasoning trace lands in ~3 minutes on either backend.