BREAKING
Google Releases Gemma 4 12B
No Encoders, Raw Inputs Direct
1
Raw 48x48 patches
↓
2
40ms waveforms
↓
3
Linear projection
↓
4
LLM backbone
0
B
Parameters
0
K
Context tokens
0
GB
Memory
Benchmark Scores
GPQA Diamond
78.8
MMLU Pro
77.2
LiveCodeBench
72
MMMU Pro
69.1
Strengths and Trade-offs
Strengths
●
Local agentic workflows
●
Audio transcription
●
30+ tokens per second
Trade-offs
●
More load on the LLM
●
Weaker fine vision detail
●
Wants more comparisons
Open Weights, Apache 2.0
AI NEWS BLITZ
Google DeepMind has launched an encoder-free open model that runs on a 16GB laptop.