Gemma 4 E4B run locally vs an OpenRouter control

Ledger entry: on-device-mobile-llm (TikTok by Eilon Grouper). User ask: "fact check it, and download it to my phone if it's a real app cause I want to try it. Also try one of the models seen in this video, let's see if it's actually usable. model: gamma 4 e4b."

Hypothesis

(1) Google AI Edge Gallery is a real, official Google app. (2) A Gemma "e4b" model exists and runs usably offline. (3) The caption's claim "free, no usage limit, prompts stay on device" holds.

Hard constraints

  • Never install anything on the user's phone or sign in with their accounts. Phone steps are prepared for the user, not executed.
  • Local model downloads are large (several GB): check free disk and RAM first; ask-first rule: if a download exceeds 8 GB total, stop and report instead.

Cases

  1. Fact check the app: official source (Google's GitHub / Play Store / App Store listing), publisher name, license, supported platforms (the video shows iPhone; the user's phone OS is unknown, so give Android and iOS). Record URLs and access dates. Pass = publisher verified as Google.
  2. Fact check the model name: "gamma 4 e4b" is the user's spelling; establish what the real name is (Gemma 3n E4B? Gemma 4?). Cite the official model card, size, license, and whether it's in the app's model list. Do not assume; if the name doesn't exist, say so.
  3. Claim check: "free", "no usage limit", "stays on device": evidence from the store listing, privacy policy and the app's open-source code if available. Mark each as verified / not verified.
  4. Run the model locally on the workstation (llama.cpp or the lab's existing runtime in .local/ / scripts/run-qwen-local.sh; Ollama is not installed). If hardware can't run it, report that with numbers, and do not force it.
  5. Usability matrix, 3 trials each, fixed prompts in source/prompts.md: short English Q&A, Hebrew Q&A (the video demo), summarize 300 words, extract JSON, 20-line code. Record tokens/s, time-to-first-token, memory, and quality (judge or checklist).
  6. Control: one small cloud model (OpenRouter) on the same prompts for a quality reference.
  7. Phone guide: write outputs/phone-install.md: exact install links and steps for the user, plus a 5-prompt script for them to try on the phone and a table to fill in (speed, battery drop, answer quality).

Deliverables

insights/RESULTS.md: verified vs not verified facts, the model's real name, workstation performance, usable-or-not verdict by task type. demo/index.html: the crux use case, offline private chat with a local model, built from the real recorded responses.