Agentopia ported to a single 12 GB consumer GPU runs 52 simulated weeks on a 4-bit 8B model
arXiv 2608.24215 (25 Aug) reimplements the Agentopia multi-agent social simulation, originally run with 100 agents over 10 simulated years on Qwen3.5-397B-A17B, on one RTX 5070 Ti with 12 GB VRAM using 4-bit Qwen3-8B-AWQ. Three stochastic runs produced 154 system-weeks (770 agent-weeks); two hit 52 weeks and one stopped at 50 on the context limit, with NO_RESPONSE fields appearing in 10.15-10.29% of activity records. The structural adaptations that made it fit are the reusable part: system-managed layered memory compression and four activity blocks per simulated day, with a memory-off ablation tying artifact production to the layered memory. The authors explicitly decline causal behavioral claims.
Source
↳ Follow the thread