Question about the custom PE-L+ backbone in SAM 3

I’m interested in the backbone architecture of SAM 3. In the paper, you mentioned using a PE-L+ encoder (450M vision parameters), but I couldn't find this specific version in the original Perception Encoder paper, which only describes a 320M "Large" version.
Could you clarify if this is a custom version optimized for SAM 3? Also, why weren't the larger models (like PE coreG or spatialG) used? Did the custom L+ version perform better than them on the SA-Co benchmark?
Thanks!

Read the SAM 3 paper and the original Perception Encoder paper, then compare their described model sizes and variants. Use the SA-Co benchmark context mentioned in the issue to determine whether PE-L+ is custom and how it compares with PE coreG or spatialG; done means documenting clear answers to these architecture and performance questions.

Written by the indexing model from the issue text.