Lucebox's engine uses hand-tuned kernels written for specific GPU and APU families rather than general-purpose frameworks, which the company reports yields several-fold throughput gains over stock stacks on the same silicon.

DFlash speculative decoding and PFlash speculative prefill, combined with 128 GB of unified memory on Strix Halo, let the box run larger models and longer context than a 24 GB consumer GPU while keeping the full stack open source.

Lucebox prices its computer as a one-time hardware purchase at a launch price of 6,499 US dollars through August 2026, rising to a standard 7,900 US dollars afterward, with global shipping included in the listed price.

Because production capacity is limited, the company uses an application and allocation flow rather than open checkout, confirming fit and delivery timing on a short call before sending a secure purchase link.

Lucebox sells a plug-and-play local AI inference computer built around an AMD Ryzen AI MAX+ 395 with 128 GB of unified memory and an AMD Radeon AI PRO R9700 32 GB GPU in an 11.97-liter aluminum chassis.

The machine ships with Ubuntu and the open-source lucebox-hub inference engine pre-installed, exposing OpenAI- and Anthropic-compatible endpoints for running large models on premises without a cloud path.