gemma4_unified¶


Model type	`gemma4_unified`
Class	`Gemma4UnifiedModel`
Task	`gemma4-unified`
Source	`models/gemma4.py`

Description¶

Unified gemma-4-12B (gemma4_unified) multimodal model.

Encoder-free counterpart to :class:Gemma4Model: it shares the gemma4 text decoder and the multimodal-fusion embedding sub-model, but replaces the SigLIP vision tower and Conformer audio tower with the lightweight encoder-free embedders (:class:_Gemma4UnifiedVisionEmbedderModel and

class:: _Gemma4UnifiedAudioEmbedderModel).

Builds a 3- or 4-model package (built by :class:~mobius.tasks.Gemma4UnifiedTask):

Always produced:

decoder: gemma4 text decoder taking inputs_embeds (and input_ids for the vision-block bidirectional mask, which it derives internally; dual head_dim, k_eq_v)
vision_encoder: raw-patch vision embedder
embedding: scaled word embedding + multimodal feature fusion

Added when config.audio is not None:

audio_encoder: raw-frame audio embedder

Registered as gemma4_unified.

Usage¶

mobius build --model <MODEL_ID> output_dir/

from mobius import build

model = build("<MODEL_ID>")