A 144M-parameter model playing Counter-Strike on my PC
I saw a video of a Jev-like model playing Minecraft, and it was insanely strong. So I thought: how would I do this for Counter-Strike?
A SourceMod plugin controls a bot, handles aim and shooting, and reads what it sees and hears. A small Python bridge describes that in plain English and asks Julia 1 what to do next: hold, peek, crouch spray, fall back, turn toward a sound.
Julia is a 144M-parameter open model that Supersonic Labs, a small lab from Brazil, just released. It doesn’t write text, it just picks an option. Everything runs on my PC, around 28 ms per decision.
In the video, it’s me and a regular bot against Julia. It’s actually good.
It also once peeked with an empty magazine. 98% confident.
Next step: writing a demo parser, pulling the data from pro player demos and fine-tuning Julia on it.