Back to AIBriefs
AnalysisAI Models

Gemma 4 12B struggles with speech attention under large system prompts

A user reports that the unified audio model in Gemma 4 12B fails to attend to speech when a large system prompt is provided, hindering one-pass voice assistant use. The model collapses ASR and LLM steps but appears sensitive to prompt length.

·
Jun 10, 6:51 AM