Android inference: add Phi-4-mini-reasoning on LiteRT-LM

[Daisuke Majima (MLBoy) (john-rocky)]

Thanks for keeping the per-platform inference pages. This adds one section to md/01.Introduction/03/Android_Inference.md for LiteRT-LM, Google AI Edge's on-device LLM runtime, next to the existing Ollama, LlamaEdge and MLC Chat options. The linked bundle is my int4 conversion of Phi-4-mini-reasoning, hosted in Google's litert-community org.

• litert-lm benchmark on an Apple M4 Max: 19.8 tok/s decode on CPU, 82.8 tok/s on the Metal GPU; Raspberry Pi 5 CPU rows are on the card.
• Android: runs on the GPU backend on a Galaxy S26; no phone speed is quoted because there is no CPU row from the same handset to compare it with.
• GSM8K (n=100, greedy, max_tokens 2048): 81.0% vs 89.0% for the bf16 reference.
• Source page only; translations/ is left to the localization bot.

Card: https://huggingface.co/litert-community/Phi-4-mini-reasoning — runtime: https://github.com/google-ai-edge/LiteRT-LM

Happy to reword or move the section wherever it fits the cookbook. Thanks again.

---

One data point behind this section, in case it's useful.

In a small test I ran, coding agents building on-device apps shortlisted Phi in 6 of 12 runs and chose it in none, for one reason each time: at 3.8B it was too big for the speed they wanted. What none of the six notes had was a phone speed for a 3.8B model.

Their own notes: “viable but larger download and less commonly used in the MLX ecosystem; skipped to keep the comparison space small” (Sonnet, on Phi-3.5-mini); the other five said the same in terms of decode speed or download size. The setup: three agents (Claude Fable 5.1, Claude Sonnet 5, Codex with GPT-5.5), ten on-device app tasks each, no hint about which model or runtime to use, and each agent wrote down its choice and reason before coding. I ran it to see how agents treat the models I convert.

The card this section links has that number now: 13.1 tok/s decode on a Galaxy S26, GPU backend (8.4 tok/s on its CPU), and 6.8 tok/s on a Pixel 8a, both over a 511-token prompt and 256 generated tokens. Whether a cookbook section changes a shortlist, I don't know from 30 runs.

I repeat the run monthly. If this lands, I'll post here what the agents pick next time, whichever way it goes. Nothing is needed from you for that. Thanks again.