What is thundermlx?

ThunderMLX is a Python project that lets two Apple Silicon Mac computers work together as one machine to run a very large AI language model called MiniMax-M3, which has 428 billion parameters. Neither Mac alone has enough memory to hold this model, so ThunderMLX splits it across a Mac Studio with 256GB of memory and a MacBook Pro with 128GB of memory, connecting them with a fast Thunderbolt 5 cable so they can share the workload as if they were a single computer. The result behaves like a private, always on AI server sitting on someone's desk. It exposes the same kind of API that OpenAI's services use, so existing coding assistants and chat tools can talk to it without special changes. The README lists a wide range of client programs that have been tested against it, including tools like Codex, Claude Code, and OpenWebUI, and describes support for live streaming of the model's step by step thinking, tool calling for AI agents, and the ability to cancel a request in progress without leaving the system in a broken state. To keep responses fast even on long conversations, the project stores recent context in a tiered cache, keeping the most active sessions in memory and moving older ones to SSD storage so they can be reloaded quickly instead of reprocessed from scratch. The README reports detailed speed measurements from real testing, describing generation speeds in the twenties of tokens per second and time to first response often under two seconds, along with a set of safeguards meant to detect and stop the model if it gets stuck repeating itself. The project also includes simple desktop commands to start, stop, and restore the whole system to a known working state, along with a live dashboard that shows request activity, cache performance, and memory health on both machines. The README documents multiple rounds of real world testing with different coding assistants to confirm the setup keeps working reliably during long agent sessions involving file edits and tool use.