MIT's AI agent society experiment reproduced on four RTX 3090 GPUs

A HAL8 team working with Apex Fusion reran MIT's SwarmWorld experiment using open-weight Qwen3.8-27B at 4-bit precision on four consumer RTX 3090 GPUs. The run took about 26 hours and 12,800 model calls; nine of 12 paired effects matched MIT's direction, and the agent society hit all 16 research milestones versus 6 for a solo agent.

TNW

Open source

MIT's AI agent society experiment reproduced on four RTX 3090 GPUs

9 of 12 paired effects matched MIT's direction; the other three fell within noise
Solo Qwen3.8-27B reached 6 of 16 milestones, the society reached all 16 (MIT: 14)
The run took ~26 hours and 12,800 model calls on four RTX 3090s
KV cache precision shifted results more than the gap between 27B and 235B models

#MIT#ApexFusion#Qwen#Nvidia

Context Language Models let AI agents edit their own working memory

Science

MIT builds robotic lab that assembles and runs its own laser experiments

Hardware

MIT's AI agent society experiment reproduced on four RTX 3090 GPUs

Context Language Models let AI agents edit their own working memory

MIT builds robotic lab that assembles and runs its own laser experiments

Four Tesla V100 GPUs on NVLink Beat M5 Ultra and DGX Spark by Up to 230% in AI Tests

Meta, MIT and UW unveil Context Language Models that manage their own context