Thunder is looking to give the idea of "running large models locally" a more powerful solution. On September 20, Thunder announced that its first AI mobile workstation, AI Master M7000, will be available for purchase tomorrow (September 21) at 10:00 AM. It claims to "tame" a 120-billion parameter model on the edge, with a final price of 19,900 RMB after combining the initial discount and national subsidies. The device covers four scenarios—creation, development, research, and confidential work—all in one scope. This means that what used to only reside in server rooms can now fit into a 2.45kg machine.

The "heart" of this device is the AMD Ryzen AI Max+395 processor, paired with 64GB of memory and a dedicated AI SSD. It comes pre-installed with the local inference engine Herdsman (Herdsman), allowing users to start working immediately upon booting up without the hassle of environment configuration. The screen is a 16-inch 2.5K panel, with an internal 99Wh battery. The chassis includes OCuLink, 2.5GbE Ethernet port, and a microSD UHS-II card reader, showcasing a layout that clearly reflects the look of a workstation.

image.png

What really enables a 120B model to run smoothly on 64GB of memory is Phison's aiDaptiv+ solution. Its approach is clever: treating the SSD as an extension of the GPU memory, temporarily borrowing weights from the solid-state drive during inference, effectively expanding the memory capacity without being noticeable. With this strategy, models like GPT-OSS-120B no longer need to rely on professional GPUs; even consumer-grade memory can handle long context throughput.