TradeMAV has introduced RAMDeck, a hardware tool designed to pool memory across multiple computers and phones to run large local AI models. The system splits neural network layers across connected Windows, macOS, Linux, and Android devices over a local network instead of requiring expensive RAM upgrades. Early backers can pre order the unit on Indiegogo starting at $109, though processing speeds vary heavily based on model architecture.
RAMDeck distributes machine learning workloads by relying on the open source llama.cpp software tool with remote computation support. The system avoids physically merging RAM sticks. Instead, each connected computer or phone stores a dedicated slice of the model weights and passes activation data across the local network during token generation. This setup allows hardware owners to load massive parameters that exceed the memory ceiling of any individual machine.
Hardware testing demonstrates mixed results depending on the chosen model architecture. TradeMAV paired a mini PC containing an RTX 3060 graphics card with an M4 Mac mini, reaching 39 tokens per second while running the Qwen3 Coder 30B model. Dense architectures cause noticeable data transfer bottlenecks across the network, dropping generation speeds down to 2 tokens per second. Pooling hardware provides the capacity to open heavy files but does not ensure fast output.
TradeMAV opened backing for the project through an Indiegogo campaign. Early bird pricing starts at $109 for initial buyers looking to bypass soaring desktop memory prices. Retail shipping timelines will depend on final crowdfunding milestones.
