Giving a game an ending, even when the user never asked, can give models an edge.
Kimi K3.1 Pro is the #1 open-weights model on our Game Dev leaderboard, carrying forward the title of its predecessor, Kimi K3, which had held it since August.
And it climbs higher: K3.1 Pro jumps five places to round out the top five, from K3's #10.
So what did reinforcement learning change in how @Kimi_Moonshot's model thinks?
We were surprised to find that the largest differences centered on game state.
Three of K3.1 Pro's four biggest gains are about keeping track of where a game stands. It thinks nearly three times as much as K3 about the progression economy (prices, money, and unlocks), more than twice as much about saving, loading, and continuing, and about 60% more about menus, the start screen, and resetting.