Lucebox pairs an RTX 3090 with a Ryzen AI MAX+ 395 and a custom inference engine tuned for it. It beats every machine at its price on tokens per second, and ships with everything pre-installed.
Key Features
Hybrid memory architecture: 128GB of LPDDR5X unified memory on the Ryzen AI MAX+ 395 holds large models, while the RTX 3090's 24GB of fast GDDR6X serves as a high-bandwidth tier. Speculative decoding across the two tiers delivers up to 10x faster inference than comparable single-tier machines.
Custom open-source inference engine: Lucebox ships with hand-tuned CUDA kernels, DFlash speculative decoding, and PFlash speculative prefill (10x faster than llama.cpp), all open source with 2,000+ GitHub stars and an active contributor community.
One-command model deployment: A single CLI pulls, configures, and serves any open model. No driver hunting, no quantization guesswork, no environment setup. Plug it in and run inference in minutes.
Pre-tuned for the exact hardware: Unlike generic builds, the entire software stack is optimized for this specific chip pairing, so you get the full performance the silicon is capable of, out of the box.
Benefits
Best tokens per second for the money: At $4,900, Lucebox outperforms every machine in its price class, including the Mac Studio and NVIDIA DGX Spark, on local LLM inference speed.
Your data never leaves your desk: Models run entirely on-device, so prompts, code, and documents stay private. No cloud provider sees your traffic.
Predictable costs instead of token bills: One upfront purchase replaces escalating cloud API spend. Run agents 24/7 without watching a usage meter.
Zero setup time: Everything arrives pre-installed and pre-tuned. You skip the weeks of kernel tweaking, driver debugging, and benchmark chasing that a DIY build demands.
