SwarmWorld Agent Society Result Reproduced on Consumer GPUs
A team working with Apex Fusion reproduced the central result of MIT’s SwarmWorld experiment on four RTX 3090 GPUs, using the 4-bit Qwen3.8-27B model and MIT’s unchanged code and commands. The roughly 26-hour run used 12,800 model calls; nine of 12 paired effects matched the original direction, while the compressed model reached all 16 research milestones in society compared with six when working alone. The result came from a single run at one population size, and larger MIT populations were not rerun.
The replication indicates that MIT’s reported collective-performance advantage did not depend on a hosted frontier model, while also highlighting that KV-cache precision can materially affect self-hosted agent benchmarks.
In August, researchers at MIT dropped a society of AI agents into a shared world with no instructions. They showed it could invent, build and maintain technology better than the same number of agents working alone. The study ran on a hosted frontier model.
A research team working with Apex Fusion has now reproduced the central experiment on four RTX 3090 gaming GPUs. Apex Fusion is a Swiss foundation building infrastructure for AI agents. It used MIT’s code and commands unchanged. The run took about 26 hours and 12,800 model calls. The paper, posted on Apex Fusion’s GitHub, and the full trace data are public.
#MIT#Apex Fusion#HAL8#SwarmWorld#Qwen3.8-27B#RTX 3090#Vector#Prime#AP3X#Agentic AI Foundation#Linux Foundation#Markus Buehler