ganesh-gemma4-e4b-v3-LiteRT

ganesh-gemma4-e4b-v3 — LiteRT builds

On-device builds of a Gemma 4 E4B model fine-tuned on Ganesha canon — shlokas, aartis, stotras, rituals, stories and Ganeshotsav 2026 dates — for Marathi, Hindi, English and Sanskrit. No retrieval: the canon is in the weights.

Which file to download

Use ganesh-v3-int8-verified.litertlm (7.6 GB).

It is the only file here that has been executed and checked. It recites Sukhkarta Dukhharta complete and correct, and answers English questions coherently.

Needs roughly 12 GB RAM. Flagship devices only.

Do not use the int4 build

model_q.litertlm produces output that begins correctly and then collapses:

सुखकर्ता दुखहर्ता, अमीर ठेवून thrombopoeto karyalayimba...

Correct opening words, then a runaway loop with characters from unrelated scripts spliced in. int4 is too aggressive for this model. int8 is not.

Two things that break Gemma 4 LiteRT exports

1. Use Gemma 4's own chat template. Gemma 4 uses <|turn> / <|channel>, not Gemma 3's <start_of_turn>. A model exported with the Gemma 3 template loads fine and then emits its own control tokens as visible text.
2. --externalize_embedder is mandatory — the exporter asserts it.

Loading on a phone

Google AI Edge Gallery (Android). Import model → from local file. There is no official iOS build.

Known gaps

Handles transactional questions ("will fasting get me a job?") by dodging rather than answering. Puranic story variants are flattened. Step-by-step ritual instructions are uneven.

Built in service of Shree Ganesh by Ravi Kadam — https://www.linkedin.com/in/ravikadam/