Wired into MiniMax H3: Jev scores each layer and picks a 1–10% attention sparsity. On an RTX 4070, 6:07 dropped to 3:34.
0 online0 visits
What are people building with Jev?
1 project · @sep_is_heim on X ↗
Wired into MiniMax H3: Jev scores each layer and picks a 1–10% attention sparsity. On an RTX 4070, 6:07 dropped to 3:34.