Components / production specification

Every part of the workstation, and why it is there.

Lucebox pairs 128 GB of unified memory with a 32 GB workstation GPU in an 11.97-litre aluminium chassis, assembled and burned in by hand. This page lists each component of the production configuration with the numbers we publish.

Lucebox workstation, top three-quarter view

01Compute platform

AMD Ryzen AI MAX+ 395 with 128 GB of unified memory

The Strix Halo platform shares 128 GB of LPDDR5X-8000 between the CPU and its 40-core integrated GPU at 256 GB/s. Models that do not fit in 32 GB of VRAM run here: long context, MoE models such as DeepSeek V4 Flash, or a second model kept resident while the R9700 serves the hot path.

Memory128 GB LPDDR5X-8000, shared by CPU and integrated GPU
Bandwidth256 GB/s
Integrated GPU40 compute units, up to 60 TFLOPS
ProcessorAMD Ryzen AI MAX+ 395, 16 cores / 32 threads
RoleLarge models, long context, and the second resident model
Black Lucebox enclosure, side and front panel view

02Dedicated GPU

AMD Radeon AI PRO R9700

The only discrete GPU in the production configuration. 32 GB of GDDR6 at 640 GB/s keeps the weights and KV cache of a 27B-class model on the card, which is where the engine’s custom kernels and DFlash speculative decoding run.

Memory32 GB GDDR6
Bandwidth640 GB/s
Compute64 compute units, up to 191 TFLOPS at FP16
Board powerUp to 300 W under sustained inference
InterfacePCIe 4.0 extension to the compute board, controlled length and strain relief
Lucebox workstation, front three-quarter view

03Enclosure

Machined aluminium chassis

Every chassis is machined from aluminium and hand-assembled in-house. At 11.97 litres it holds a full-length workstation GPU, the compute board, and an SFX power supply in a box that fits a shelf or a rack area.

EnclosureMachined aluminium, hand-assembled in-house
Dimensions340 × 110 × 320 mm (13.4 × 4.3 × 12.6 in)
Volume11.97 L (approximately 730 in³)
Weight6 kg (13.2 lb)
FinishBlack
Thermal camera time-lapse: the processor package warms from 23 to 42 °C as inference starts

04Cooling and thermals

Two air paths, exhaust out the back

A 120 mm fan cools the compute region and the R9700 runs its own cooler, each with an independent airflow path that exits through the rear. Quiet at idle; the GPU becomes audible during sustained inference, which is why most people keep Lucebox out of earshot and reach it over the network.

CoolingAir-cooled, tuned for low noise under sustained load
Compute region120 mm top fan over the compute board
GPUR9700 cooler on its own airflow path
ExhaustHot air leaves through the back of the chassis
AcousticsQuiet at idle, GPU cooling audible under sustained inference

05Full specifications

Power, storage, connectivity, system.

Power

Output
1,000 W, SFX form factor
Efficiency
80 PLUS Platinum
Draw
~500 W under full inference load, ~40 W idle
Mains input
IEC C14 inlet, plugs into a standard wall outlet
Power button
Momentary chassis switch wired to the board header

Storage

Capacity
2 TB
Interface
NVMe 1.4, M.2
Throughput
Up to 7,400 MB/s read, 6,600 MB/s write
Contents
Ubuntu Server, Lucebox engine, tuned model profiles, your own GGUF weights

Connectivity

Ports
2× USB4, 2× USB-A, HDMI 2.1, 2× DisplayPort 2.1, 5 GbE Ethernet, plus 4 display outputs on the GPU
Wireless
Wi-Fi 7 and Bluetooth, built in
Network
5 GbE Ethernet or Wi-Fi 7
Setup
Bluetooth pairing from a phone or laptop, no monitor needed

System

Software
Ubuntu Server, headless, engine and models pre-installed, OpenAI and Anthropic compatible API
Warranty
3 years, parts and labor, 72-hour burn-in before shipping
Access
Full root access over SSH, it is your machine
Offline
Runs fully offline once a model is loaded; internet only for updates

Order

The whole workstation, from $5,999.

Both tiers ship the same hardware listed above. The Business tier adds a year of Lucebox Engine Pro. Worldwide shipping and taxes are included, and checkout runs on Stripe.