One M5 Ultra 96GB: the 27B-and-under sweet spot with 100B-class quantized guests.
TECHNICAL DIFFICULTY: 1 OF 4 ยท PLUG IN AND GO
How hard is this to assemble? one box, one cable: it ships whole, plug it in
For most people actually running local AI daily, this is the honest machine. 96GB of unified memory runs every small decision model natively, the 27B multimodal class at full precision, and the 100B+ class in 4-bit when you are patient. It is silent, it uses less power than the monitor next to it, and it costs a quarter of the flagship build. It will not hold the biggest models at long context; that is a 512GB conversation.
WHO THIS BUILD IS FOR
You use local AI every day and want to stop thinking about the hardware. Writers, developers, and small teams who want a private coding assistant, document search, and decision models running around the clock. If your work touches client data, this tier is where local stops being a demo.
WHAT IT WILL NOT RUN
DeepSeek V4.1 Flash: 552B total parameters is 138GB of weights even at 2-bit. GLM 5.3 Flash runs at 2-bit with short context only. Video generation wants an NVIDIA card, which is why the 50k build carries a 5090.
POWER, NOISE, AND THE ROOM IT LIVES IN
About 180W under load and effectively silent at desk distance. One cable, no UPS required.
NEXT: THE $12k TIER
Open the case. This is where local AI becomes a hobby with a parts list: two GPUs, a power plan, and the fastest single-machine inference you have ever seen. New models land on CUDA stacks first, and this build is standing on the loading dock when they arrive.
See the $12k tier →