LuceBox Hub Achieves 207 tok/s Running Qwen3.5-27B on RTX 3090 — Show HN with 164pts

GitHub·medium signal

LuceBox Hub, a new inference optimization project, demonstrated 207 tokens/second running Qwen3.5-27B on a single RTX 3090, generating 164 points and 45 comments on HN. For context, this is a 27B parameter model running at consumer-grade speed on previous-generation hardware. The project signals that the gap between cloud and local inference continues to narrow — builders can now run competitive open-source models at interactive speeds without enterprise GPU hardware, making local-first agent architectures increasingly viable for solo developers.