Google's Gemma 4 12B debuts encoder-free multimodal architecture

Google has released Gemma 4 12B, a new multimodal model that notably omits traditional specialized encoders for vision and audio. Instead, it processes these inputs directly through its decoder-only transformer backbone, aiming to reduce latency and simplify the architecture. This 12-billion parameter model is designed to run on consumer hardware with 16GB of VRAM, filling a gap in the Gemma 4 lineup for capable local agentic systems. AI

IMPACT This novel architecture could reduce latency and simplify multimodal AI development for local agentic systems.

RANK_REASON New model release from a major AI lab with a novel architectural approach. [lever_c_demoted from frontier_release: ic=1 ai=1.0]

Read on Towards AI →

• Gemma 4
• Gemma 4 12B
• Google

• model release
• product

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →