cong_ml runs 8-agent research swarm on a single 48GB M4 Pro with one local 40B model

congml demonstrated a research agent swarm running entirely locally on a single Mac: 8 agents share one 40B Flint model via llama.cpp on a 48GB M4 Pro, writing code, testing ideas, and iterating on the best results — the user only supplies a problem and a score to beat. In practice, batched concurrent inference nearly doubled total throughput, with the 8 agents producing 2.3x more effective experiments per hour than a single agent.

Confirmed

Hardware and model: a 48GB M4 Pro Mac running one local 40B Flint model in llama.cpp, with 8 agents sharing the same weights.
Concurrency: a single llama.cpp instance with server-side batching handles 8 concurrent requests; on the M4 Pro, total token throughput across 8 concurrent requests is 1.9x that of a single request — each agent writes more slowly, but combined output nearly doubles.
Collaboration: agents share the "current best program"; once one agent breaks the record, the others start the next round from the new best. The author noticed that agents starting from the same program tend to try the same directions, so each agent takes turns receiving one of four different instructions — refine a part, switch algorithms, and so on — to avoid homogenization.
Results: across three tasks (circle packing, job scheduling, and bike demand forecasting), 1 vs. 8 agents each ran six hours per task, twice per task: the 8 agents averaged 2.3x more effective experiments per hour, and charts show the final solutions were also better (e.g., circle packing).

Why it matters

This shows that multi-agent collaborative research is feasible on a consumer Mac without any cloud APIs: multiple agents share one set of model weights and use batched inference to improve hardware utilization, offering a reference path for local, low-cost research automation.

2026-10-27 ~ 2026-10-27 · 7 related posts

Primary sources

Running an 8-agent research swarm on a local 40B model with a 48GB M4 Pro — cong_ml · 2026-10-27 23:43 (UTC+08:00)
One llama.cpp instance serves 8 agents with 1.9x token throughput on M4 Pro — cong_ml · 2026-10-27 23:43 (UTC+08:00)
8 agents beat 1 by 2.3x in valid experiments per hour, six-hour local swarm benchmarks — cong_ml · 2026-10-27 23:43 (UTC+08:00)

[source] Running an 8-agent research swarm on a local 40B model with a 48GB M4 Pro — cong_ml · 2026-10-27
Running an 8-agent research swarm locally on a 48GB M4 Pro with one 40B Flint model — cong_ml · 2026-10-27
[source] One llama.cpp instance serves 8 agents with 1.9x token throughput on M4 Pro — cong_ml · 2026-10-27
How the swarm collaborates: shared best program plus rotating instruction variety — cong_ml · 2026-10-27
[source] 8 agents beat 1 by 2.3x in valid experiments per hour, six-hour local swarm benchmarks — cong_ml · 2026-10-27
Run a Research Swarm on a Mac: 8 Agents Share One Local 40B Model on 48GB M4 Pro — cong_ml · 2026-10-27
Running an 8-agent research swarm with a local 40B model on one 48GB M4 Pro — cong_ml · 2026-10-27