Four models play the same Doom level, Jev leads with 5.63 average kills
A Python adapter feeds enemies, health and ammo to the models, which only choose move left, move right or fire; across eight seeds Jev averages 5.63 kills, Qwen 3.5 4B 3.63, and Laya and ModernCE 1.25 each.
HOLY SH*T! Someone gave the same Doom level to four AI models and let them fight. Jev wasn’t shown pixels and received no Doom-specific training. A Python adapter described the visible enemies, health and ammo, then the models had to choose one action: move left, move right or fire. Across eight seeds, the results looked like this: > Jev: 5.63 average kills > Qwen 3.5 4B: 3.63 > Laya: 1.25 > ModernCE: 1.25 Jev made roughly five decisions per second while the game continued running. This is a much better explanation of Jev than another benchmark table. It doesn’t need to write a strategy, narrate its reasoning or roleplay as a Doom player. It reads the current state, returns a probability for each valid action and lets ordinary code execute the decision. See enemy -> Measure risk -> Choose action -> Repeat. That same loop can route an agent, approve a tool call, detect failure or decide when an expensive LLM is actually needed. Jev isn’t interesting because it can play Doom. It’s interesting because software can finally ask AI for a decision without paying it to write an essay first. I mapped 12 real decision systems inside the article below.
This post quotes another post: x.com/xmyttle/status/2103904144289390899