Together AI Adds Ternary Bonsai 27B for Phones
A 27B multimodal model that runs locally instead of only in the cloud
@togethercompute4 days ago
TL;DR:
- Together AI now hosts PrismML's Ternary Bonsai 27B
- Ternary weights cut memory and compute needs
- More AI could move to local devices if it works well
Together AI is now hosting PrismML’s Ternary Bonsai 27B. This 27B model works on phones and edge devices, not just in the cloud.
It uses ternary weights – basically low-precision numbers for the parameters. That means less memory and less compute. If it performs well, we'll see more AI running locally, which could mean faster responses and lower costs. Putting it on Together gives devs a chance to try it out before committing to on-device deployment.
Things are moving away from just making models bigger. Now it's about clever compression to get better efficiency.