Fifteen litres of machined aluminium around an RTX PRO 6000 Blackwell. Ninety-six gigabytes of VRAM in your study - not in someone else's data centre.
MX-15 · RTX PRO 6000 Blackwell · 15.3 litres · see all of it
The whole machine, laid out. Five layers, two panels, one hex key.
Eleven, fifteen or nineteen litres.
Open at both ends, because the card inside it draws 600 watts.
Set up for the models you already run.
Tell us the model you want to run. We load it on your exact machine, measure the tokens per second and send you the number before you pay.
Arrives with drivers, Ollama and a local chat interface on your own network. Plug it in, open a browser, talk to it.
34 dBA under full load. It sits on the desk beside you, not in a cupboard down the hall.
Fifteen litres. Ninety-six gigabytes. No cupboard required.
The frame is cut from 6082-T6 plate — one piece carrying the board, the card and the cooler. Both end walls are perforated on a 3 mm grid across 42% of their area and the long wall is tempered glass, so a 600-watt card pulls cool air straight through the machine rather than sipping it through a slot. Everything above the card is clearance for its power cable — 20 mm of it, which is why the machine is 246 mm tall and not 226.
Anodised, not painted, with stainless fasteners and a filtered intake you can rinse in the sink. Every figure below is measured on the machine that ships, not on a prototype.
Four decisions. Every one of them a part you could buy yourself.
{{ configIntro }}
Eleven, fifteen or nineteen litres.
The same frame and the same perforated walls in three volumes, sized to the card inside — RTX PRO 4500, 5000 or 6000 Blackwell. The 6000 is 304.8 mm long and wants 20 mm of cable clearance above it — that is what sets the width and height of the MX-15. Pick the one that holds your model in memory and nothing larger.
Every machine ships with its burn-in log, its dimensioned drawing and a build sheet naming every part by revision. Swapping the card yourself in two years is not a warranty event.
32 GB of GDDR7 with ECC on a Ryzen 9 9950X. The 4500 is 267 mm long, so the whole machine comes in at eleven litres. Runs a 32B model at 4-bit and fits under a monitor arm.
96 GB of GDDR7 with ECC, flow-through, exhausting through an open lid. Holds a 120B-class model resident and answers all night. The one most people buy.
The same 96 GB card on a micro-ATX board — four memory slots, a second full-length slot, and a full-height Noctua NH-U9S where there is finally room for one.
Measured. Logged. Shipped.
Name the model you want to run at home — Llama, Qwen, DeepSeek, gpt-oss. We load it on the exact machine you specified, measure the tokens per second and put the number in the build sheet before you pay.
Tell us how you like to work and the machine arrives with it installed, pinned and burned in — not a fresh Ubuntu and a link to a wiki.
{{ workflowBody }}
Tokens per second at your context length, on your configuration, written into the build sheet before you pay for it.
Mini-ITX, SFX-L, a standard card on a standard riser. Nothing inside is ours alone, so nothing inside is ours to hold you to.
Drivers, a model server and a chat interface on your own network, pinned to versions tested together. Plug in, open a browser, ask it something.
“It replaced my API bill and it is quieter than the fridge. I forget it is running until I put a hand on the lid.”
Thermal results. Part revisions. What we learned.
One email a month from the people cutting the parts — thermal results, part revisions, what broke. No launch countdowns, no discount codes.
Three years on the machine, ten on the chassis. Opening it to add memory or swap a card does not void anything.
Every machine ships with its own dimensioned drawing, torque table and burn-in log. Replacements are a PDF away.
Support is the workshop, not a queue. Mail arrives at the bench where the machines are assembled, in Singapore.