BREAKING
Google releases Gemma 4 12B
0
B
Parameters (Dense)
0
K
Max context tokens
0
GB
RAM/VRAM with quantization
Encoder-free design
1
Text, image, audio in
↓
2
Linear projection
↓
3
Unified model
↓
4
Lower latency
Benchmark scores
GPQA Diamond
78.8
LiveCode Bench
72
BBEH
53
Praise vs caveats
Strengths
Praise
●
Local coding agent on Mac
●
256K context in 16GB
●
Smooth video and audio
Caveats
Concerns
●
44 tokens/s on M5 Pro
●
Trails Qwen 3.6 family
●
Narrower knowledge breadth
Local AI fills the middle
AI NEWS BLITZ
Google DeepMind has launched Gemma 4 12B, an open multimodal model that runs on a laptop.