Latest in AI

Luce Spark: a 35B MoE on a 16 GB GPU, without the offload tax★ 72

Luce Spark is an open-source MoE offload system for running 33B-35B A3B models on 16GB-class GPUs. It keeps frequently routed experts on GPU, stores the long tail in system RAM, and swaps cold experts through a bounded async cache. The author reports 13.3 GiB for Qwen3.6 35B-A3B and about 100 tok/s with Spark optimizations, but notes real 16GB GPU testing is still missing.