BREAKING
Apple Eyes PrismML 1-Bit Tech
0B
parameters
0GB
memory use
0x
size reduction
Memory Footprint by Model
Bonsai 8B1.15
Bonsai 4B0.5
Bonsai 1.7B0.24
0
benchmark avg
0
GSM8K
0
HumanEval+
Native 1-Bit vs Conventional AI
PrismML 1-bitnative
Weights are +1 or -1
No high-precision layers
Apache 2.0, runs on MLX
ConventionalFP16
8B model needs ~16GB
Hard to fit on edge
Often needs the cloud
On-Device AI Talks in Play
AI NEWS BLITZ
Apple is in talks with startup PrismML to run big AI models directly on iPhones.