BREAKING
PrismML Unveils Bonsai 27B
0
GB
1-bit build
0
GB
ternary build
Sharp Cut From Conventional Size
FP16
54
4-bit
18
Ternary
5.9
1-bit
3.9
Benchmark Retention vs Baseline
FP16 base
85
Ternary
80.5
1-bit
76.1
Strengths and the Caveats
Holds Up
●
Math tasks
●
Coding tasks
●
Apache 2.0 license
Weaker
●
Agentic tool-calling
●
Vision tasks
●
Needs custom kernels
On-Device AI Turning Point?
AI NEWS BLITZ
PrismML says its new 27-billion-parameter model can run right on an iPhone.