No One Asked for an Ending; Kimi K3.1 Pro Tops the Open-Weight Game Leaderboard by Finishing the Game

DesignArena’s Game Dev leaderboard has been updated: Kimi K3.1 Pro is now No. 1 among open-weight models and No. 5 overall. It replaces Kimi K3, which had held that position since August. K3.1 Pro rises five places: K3 was No. 10, and K3.1 Pro is No. 5.

DesignArena is a crowdsourced benchmark. Users provide prompts to generate websites, games, images, and more, then vote pairwise on model outputs. This leaderboard looks at generated games; coding ability is only part of it.

DesignArena says the biggest post-RL changes are concentrated in game state. Three of K3.1 Pro’s four largest improvements relate to “how far the game has progressed.” Its reasoning about progression economy—including prices, money, and unlocks—is nearly three times that of K3; its reasoning about save, load, and continue is more than twice as much; and its reasoning about menus, start screens, and resets is about 60% higher.

This shows up in the results. The two models start games at similar rates: K3.1 Pro at 94%, K3 at 91%. But 62% of K3.1 Pro’s games can be completed or finished, versus 40% for K3.

More counterintuitively, 70% of prompts never asked for win conditions, levels, goals, or quests. Under these prompts, 53% of K3.1 Pro’s games can still be completed, versus 28% for K3. When prompts explicitly ask for an ending, the figures are 84% versus 71%. When the same users prompt the two models separately, the share of completable games is 73% for K3.1 Pro and 30% for K3.

Users reward this behavior, too. With opponents fixed at their respective leaderboard ratings for comparison, within the same model, a game that can be completed earns about 83 rating points more in head-to-head voting than one that cannot.

Some users said in replies that this niche scoring advantage for “finishing the game even when no one asked” is odd, and that competition among open-weight models in game development is starting to heat up. Hana Kaito AI noted that RL improves game-state awareness, not just coding. MASA asked more directly: will prioritizing state hurt core gameplay depth? The leaderboard cannot answer that question, but it is worth including in the next round of evaluation.

K3.1 Pro’s changes show that a model filling in goals the user did not state can itself become a product capability. An evaluation that rewards completion will push models to learn how to wrap things up. The side effects also need watching: generated games may lean more toward completable structures, and gameplay depth may not rise in tandem.

Source: DesignArena’s X post https://x.com/DesignArena/status/2117315418528575649 , interactive version https://designarena.ai/blog/how-rl-improved-kimis-games , playable entry https://DesignArena.ai .