The Luce-Org team achieved 207 tokens per second when running Qwen3.5-27B on a single RTX 3090 GPU. The result demonstrates significant performance gains for the 27-billion parameter model on consumer hardware.
Luce-Org's lucebox-hub project published benchmark results showing Qwen3.5-27B reaching 207 tok/s inference speed on an NVIDIA RTX 3090 graphics card. The throughput metric indicates practical improvements for deploying mid-size language models on widely available consumer GPUs.
Qwen3.5-27B sits in the sweet spot between capability and resource efficiency, making it viable for developers working with constrained hardware budgets. The RTX 3090, while not entry-level, remains popular among researchers and small teams.
The post generated significant discussion on Hacker News with 127 upvotes and 31 comments, suggesting community interest in real-world performance metrics for open-source models. Performance benchmarks like these help developers make informed decisions about model selection for production deployments.
Details on optimization techniques used to achieve these speeds are available in the lucebox-hub repository.
Xreal has launched its Air 01+ augmented reality glasses at $299. The lightweight device targets gamers and video viewers seeking portable large-screen experiences.
Motorola's 2026 Razr and Razr Plus flip phones cost $100 more than their predecessors while offering few meaningful improvements, exemplifying the smartphone industry's shrinkflation trend.
SK Hynix CEO Kwak Noh-Jung forecasts the memory industry will face its worst-ever supply shortage in 2027, with demand outpacing supply extending beyond 2030.
QuadRF, a radio frequency detection device, can identify drone activity and visualize WiFi signals penetrating through physical barriers. The technology demonstrates RF sensing capabilities that extend beyond conventional wireless detection methods.