Apex Fusion Shows MIT's AI Agent Societies Do Not Need a Frontier Model to Work
Apex Fusion Foundation reran MIT's SwarmWorld experiment with 50 agents on four consumer RTX 3090 GPUs in about 26 hours and 12,800 model calls, with MIT's code unchanged. Nine of twelve paired effects matched MIT's direction, including every portfolio-resilience gain, which was larger in the replication. A single Qwen3.8-27B agent reached 6 of 16 research milestones against 7 to 10 for MIT's model, while the society reached 16 of 16 against 14. A hidden KV-cache precision setting changed results more than the gap between a 27B and a 235B model. The swarm also ran on Vector, Apex Fusion's sett
Discussion
Rejoignez la communauté pour réagir et poser vos questions.
Aucun message pour l'instant. Lancez la discussion !