My experience so far is that Jev is extremely fast and capable of avoiding immediate risks, but Luna makes better decisions when the consequences extend several moves into the future.

My experience so far is that Jev is extremely fast and capable of avoiding immediate risks, but Luna makes better decisions when the consequences extend several moves into the future. I tested different prompt calibrations, from safety focused to more apple focused, while always giving both models the same prompt and context. Jev either became too cautious or pursued apples without anticipating later traps as consistently, whereas Luna more often balanced progress with the long-term shape of the snake. It was not perfect, but it appeared to reason further ahead, sometimes taking several seconds and using more reasoning tokens on difficult moves. This is only my experience from these experiments, but Jev’s long-horizon judgment still seems behind Luna and the expected Terra-level performance, despite its impressive speed. For a first-generation system, though, it is extremely impressive, and I’m excited to see how it improves as it becomes more refined for long-horizon decision tasks.