Lucebox explores speculative inference on consumer GPUs

Published on May 9, 2026, this entry points to Lucebox, described as a speculative inference server intended for heterogeneous hardware and consumer GPUs. The post claims that Qwen3.6-27B reaches 120–200 tokens per second on an RTX 3090; this is the post’s claim, not an independent measurement provided in the entry.

The repository is presented as material for engineers interested in exploring speculative inference on consumer GPUs. To consult and verify the result, review the original post and repository, checking for details about configuration, workload, and measurement method before comparing the reported rate with your own tests.