Can a cluster of weak, mismatched hardware run a 120B parameter model? I connected 6 completely different devices — two Windows machines (one with an RTX 3060, one an old 12GB laptop), three different Macs (Intel CPU-only, M3 Metal, and a Mac mini), and an Android phone — and used RAMDeck to pool them into one system.
In this video:
– Touring the full 6-device cluster and how it's networked (Ethernet + Wi-Fi mix)
– Loading a 4-bit quantized 120B GPT model (~60GB across two shards) with the mini PC as primary
– Watching RAMDeck's waterfill logic prioritize the RTX GPU, then the Metal devices, then CPU fallback
– Chatting live with the model once it comes online
One critical lesson: why your load timeout needs to scale dynamically with model size (mine took ~11 minutes)
RAMDeck — Pool Your RAM. Run Bigger AI.
Run and train models too big for any single machine, using the PCs, Macs, mini PCs and phones you already own.
Leave a Reply