BREAKING
Google Releases Gemma 4 12B Unified
Encoders Removed From the Stack
1
Raw image, audio, video
↓
2
Lightweight projection
↓
3
LLM embedding space
0
B
Dense params
0
K
Context length
0
GB
Quantized RAM
Memory: Quantized vs 16-bit
Quantized
16
16-bit
26.7
Strengths and Open Questions
Advantages
●
Runs on 16GB Mac Mini
●
Lower latency, easy fine-tuning
●
Apache 2.0 license
Cautions
●
Quantization affects performance
●
Complex vision and audio unproven
●
Beats 26B case-by-case
Local Multimodal AI Advances
AI NEWS BLITZ
Google has launched Gemma 4 12B Unified, its first mid-sized encoder-free multimodal model.