Part 3: 5 #AIAgents on one 16GB MacBook.
Qwen3-8B agents design, code, QA, backtest and paper-trade a strategy.
Wait per agent turn:
Stock mlx_lm.server: 15.7s
purgekv-server: 7.0s
Their memory goes to SSD and back while tools run.
#LocalLLM#MLX
https://t.co/4Oqpf7qRCH
I just got Devin Max for a month for free and decided why don't I make my website better, its something i have been wanting to do for a while and the localhosts has had too many ideas so I decided its about time to be a #meatproxy
check it out and lmk https://t.co/qbjBdkbZFx
@Hoagie78Ai ~3 GB for the full 5-agent run (19 turns, Qwen3-8B, about 147 KB per token of KV). --kv-bits 8 halves that.
Your question made me check, and I found Qwen3 was rewriting the whole chat on every save. Fixed: ~3.4x fewer writes in my tests. I’ll post a measured number soon
@cubicles_cloud@KalinJunYAI@cubicles_cloud Also the demo process is meant to show a queue of agents due to it needing the research done before we code and then before QA etc. If I run them at once there’s no ability to trade bc there’s no trading algorithm. Hope this helps!
@cubicles_cloud Yup, what it changes is what the queue costs. 15.7s per turn is stock mlx-lm recomputing each agent’s memory after it gets pushed out of RAM. purgekv brings it back from SSD, and the wait drops to 7.0s.
@Hoagie78Ai So reads don't wear the SSD its only the writes, due to this I decided to make it so that only new tokens are written to the SSD. This means that a majority of the operations don't wear it however few do but they're minimized as much as possible