Jev has started overseeing Attention, cutting MiniMax H3 generation time by 41.7%.

Beating AI News Flash: Japanese developer Kamimoto has integrated Jev into the MiniMax H3 video generation pipeline. It selects how much Attention to retain from four levels—1%, 3%, 5%, and 10%—based on the operating status of each layer, and then Sparse Attention determines which specific computations can be skipped.

On an RTX 4070 12GB, generating the same roughly 5-second video takes 6 minutes 07 seconds with the original 4-step approach, but drops to 3 minutes 34 seconds after adding Sparse Attention + Jev dynamic control, cutting total time by 41.7%.

However, this 41.7% cannot all be attributed to Jev. Sparse Attention itself can already speed things up by computing less Attention, and fixed retention at 5% or 10% would also be faster. What Jev does is dynamic adjustment: compute a bit more for important layers and a bit less for unimportant ones.

At present, it cannot yet be proven that this dynamic adjustment definitely produces better image quality than a fixed ratio, and the author also explicitly states that quality has not been fully validated. The experimental code has already been open-sourced.