RAMDeck wants to pool old devices for bigger local AI models
TradeMAV’s $109 RAMDeck campaign uses llama.cpp’s existing RPC tools to pool memory across old PCs, Macs and phones — with speed trade-offs.
TradeMAV is crowdfunding RAMDeck, a small hub designed to let Windows PCs, Macs, Linux machines and Android phones contribute memory to a shared local AI workload. The project is listed at $109 for early backers on Indiegogo, with the campaign due to close on 8 November 2026. 1
What RAMDeck is actually doing
RAMDeck does not merge the physical memory in a pile of computers into one giant hardware pool. Instead, its software divides a model’s layers between participating devices, keeps each portion in local memory and passes intermediate results across the network as the model generates tokens. TradeMAV says the system can discover nodes automatically, place layers across available hardware and expose an OpenAI-compatible API. 2 3
The underlying technology is less mysterious than the product pitch suggests. TradeMAV’s public code uses llama.cpp’s RPC backend, which already allows compatible machines to contribute to inference without buying a RAMDeck. The open-source exo project offers another route for distributing AI workloads across computers, phones and single-board systems. 3 4
Capacity comes before speed
TradeMAV’s own figures show the trade-off clearly. Mixture-of-experts models can achieve useful speeds when the active parameter count is relatively low, but dense models become dramatically slower as more devices are added. One 27B dense model managed roughly 1.92 tokens per second across four devices, while a 120B model took several minutes to initialise and produced only a few tokens per second in the reported tests. 2
That makes RAMDeck a capacity play rather than a performance replacement for a well-equipped GPU. It could allow a model that otherwise fails for lack of memory to run at all, but network traffic and the slowest device in the cluster still shape the experience. TradeMAV’s own guide acknowledges that pooled memory trades speed for capacity. 2
A crowdfunded convenience layer
The official RAMDeck material does not yet list basic hardware details such as the processor, RAM, ports or power draw. What the project is selling is the convenience layer: device discovery, workload placement, recovery when a node drops out and a dashboard for managing the cluster. The free software underneath means technically confident users can try the concept before deciding whether the hub’s simpler setup is worth paying for. 2 3
Our opinion
RAMDeck is an interesting answer to a real problem, particularly while memory prices make large local models harder to run. But it is important not to confuse “fits in memory” with “feels fast”. The project’s strongest case is convenience for tinkerers who already own several mismatched machines, not a cheap substitute for a powerful GPU or a tightly integrated workstation.
Because this is a crowdfunding campaign, the hardware remains a promise rather than a finished product. The open-source route is the sensible way to test the idea today; RAMDeck’s challenge is proving that its orchestration is convenient enough, reliable enough and transparent enough to justify the $109 early-backer price.