System76's Thelio Mira AI Packs 192GB of GPU Memory Into a Linux Workstation
System76 is now selling the Thelio Mira AI, a Linux workstation that carries as much as 192GB of graphics memory for running large models locally.
System76 has started taking orders for the Thelio Mira AI, a Linux desktop built for running large language models on hardware you own rather than renting GPU time from a cloud provider. The headline figure is its memory ceiling: up to 192GB of graphics memory, roughly eight times what a top consumer card offers. Orders are open on the company's website now.
That number matters because VRAM, not system RAM, decides which models a single machine can load. Most consumer GPUs ship with 16GB to 24GB, so anything past a mid-size open-weights model has to be split across multiple machines or quantized until output quality drops. At 192GB, the Mira AI can hold far larger weights resident in one box, putting it in the same range as small server nodes rather than typical desktops.
System76 has not published a full component list. The product page states the memory figure but leaves the exact GPU arrangement, pricing, and power specifications unclear, which makes direct comparisons difficult for buyers weighing it against comparable workstations. Those details will decide whether the machine is a practical purchase or a spec-sheet talking point.
The Colorado vendor has sold Linux laptops, desktops, and servers running Pop!_OS, its Ubuntu-derived distribution, for years, and it controls the firmware on its hardware. That vertical integration is the pitch: a buyer gets a machine tuned for Linux out of the box instead of assembling parts and chasing driver compatibility. System76 has pushed further into AI hardware as demand for local inference has grown among developers, researchers, and teams that cannot send data off-site.
The Mira AI lands in a crowded field. Workstation builders and the major GPU vendors all sell high-memory systems, and cloud providers argue that renting compute beats owning it for all but the steadiest workloads. System76 is betting on a slice of buyers who want the hardware on their desk, under their control, with an operating system they did not have to fight.
The real test comes after delivery. Independent benchmarks will show how much of that 192GB common training and inference stacks can actually address, and how the system handles sustained load, noise, and power draw. Until those numbers are in, the Mira AI is a memory claim, not a proven platform.