IT之家 · 智能时代T2·72 pts
- 7B total and 1B active parameters using Differential-TPA plus Mamba-3 hybrid attention
- Open sourced under Apache 2.0, runs on 8GB VRAM and hits 160 TPS on an RTX A4500
- Training cost about EUR 7 million; Olala 2.0 is planned to exceed 100B parameters