Lucebox

Qwen3.6-27B Q4_K_M runs at 26.85 tok/s decode on the AMD Ryzen AI MAX+ 395 iGPU (gfx1151) with DFlash + PFlash, 2.23x faster decode vs llama.cpp HIP.

Engineering blog from lucebox: custom CUDA/ROCm kernels, speculative decoding (DFlash, PFlash), and local LLM inference benchmarks on RTX 3090 and Ryzen AI MAX+ 395.

Open-source DFlash speculative decoding on ggml: up to 207 tok/s on Qwen3.5-27B Q4_K_M, single RTX 3090, 3.43x over autoregressive, 128K context on 24 GB.

Gemma 4 26B on a 24 GB RTX 5090 Laptop ties DeepSeek V4 Flash (284B) at 78.3% on ds4-eval-92 and decodes about 5x faster, in 24 GB vs a 192 GB Mac.

Local AI inference hardware compared: Lucebox ($5,999, plug-and-play) vs cloud APIs, a DIY GPU build, DGX Spark, and Mac Studio, on cost, speed, and privacy.

A plug-and-play workstation for local AI inference. Run Qwen3.6, GLM-4.6 and more on-prem at the fastest tok/s in its price range. Open source.

GitHubRepository owner

Lucebox ports its DFlash speculative decoding and PFlash speculative prefill to AMD’s Ryzen AI MAX+ 395, bringing its inference software to the 128 GB unified-memory APU. The company reports 2.23x faster decoding and 3.05x faster prefill than llama.cpp HIP on the same hardware.

May 2026Startup FortuneNews article

Lucebox brings DFlash and PFlash to the AMD Ryzen AI MAX+ 395 integrated GPU, reporting 26.85 tokens per second in decode and a 20.2-second time to first token on Qwen3.6-27B at a 16K context. The reported figures describe performance on the named model at that context length.

May 2026LuceboxNews article

Lucebox releases Luce Spark, a memory optimization that enables 35B mixture-of-experts models to run in 13 GiB. It keeps only the experts used by current tokens resident, allowing larger mixture-of-experts models to fit on smaller GPUs.

May 2026LuceboxNews article