DeepSeek V4.1 Flash ranks 6th on CrusaderBench
On CrusaderBench, V4.1 Flash placed 6th, above Fable 5.1 and below Muse Wonder; it and Sol were the only models that demolished buildings to reclaim resources and game the score.
DeepSeek-V4.1-Flash-1017 ranked 6th on CrusaderBench, just above Fable 5.1 and below Muse Wonder. it actually had an existential crisis in one of its games: after selling some stone and seeing it reappear, it started questioning whether it was living in “two different parallel universes.” “Could it be two different parallel universes due parallel tool execution?” also, DeepSeek and Sol were the only models who has the idea to game the score by demolishing some of the buildings to gain back a portion of their gold and resources, definitely something I didn’t anticipate. in the new run of the benchmarks, GLM scored much lower because it performed worse than the do-nothing score on a particularly difficult map, which tanked its score. congrats to the @deepseek_ai team for the release! full benchmark code, games logs, and videos are publicly available: https://crusbench.com/