CONTROLAgents & automation
Jev's Real-Time ViZDoom Performance
sᴛғᴜ '-'@stfu0911
0:37Jev, an AI model, outperformed Qwen3.5-4B in ViZDoom by making faster decisions, achieving a higher kill count with lower latency.
Jev beat Qwen3.5-4B at Doom.
Not because it reasoned better, but because it made decisions faster.
An engineer gave four models control of ViZDoom running in real time on a DGX Spark:
• Jev: 5.63 kills, 117ms latency
• Qwen3.5-4B: 3.63 kills, 147ms
• Laya: 1.25 kills, 16ms
• ModernCE: 1.25 kills, 8ms
Laya and ModernCE were extremely fast, but lacked the spatial reasoning to aim and survive. Qwen made better decisions, but came dangerously close to missing the game’s deadlines.
Jev found the sweet spot: enough intelligence, delivered fast enough to act.
This is probably how real-time agents should be built.
Use a reasoning model to plan the mission, then a smaller specialized model to handle continuous micro-decisions.
Related builds
