gemma3n

Model type	gemma3n
Class	Gemma3nMultiModalModel
Task	gemma3n
Source	models/gemma3n.py

Description

Gemma 3n image + audio + text model (3- or 4-model split).

Always produced:

• decoder — text decoder on inputs_embeds + per_layer_inputs

• vision_encoder — MobileNet-V5 tower + embed_vision soft path

• embedding — token embedding, hard placeholder embedding, feature fusion, and the per-layer input tables

Added when config.audio is not None:

• audio_encoder — USM Conformer tower + embed_audio

Registered as gemma3n; the text-only gemma3n_text key keeps mapping to :class:Gemma3nCausalLMModel.

Usage

mobius build --model <MODEL_ID> output_dir/

from mobius import build

model = build("<MODEL_ID>")