Janas runs a 48 GB AI model on a 32 GB laptop at 37 tok/s

Developer Maurizio Cammalleri released Janas, a C engine that streams MoE weights from an NVMe SSD via O_DIRECT, caches hot experts and uses speculative drafting. In its own benchmarks, Qwen3-Next-80B (48.4 GB) hits 37 tok/s on a 32 GB laptop versus 2.8 tok/s for llama.cpp.

• Qwen3-Next-80B: 62 tok/s prompt read and 24 tok/s write without drafts on a Core Ultra 9 185H
• Speculative drafting raised write speed from 24 to 37 tok/s
• Janas beat llama.cpp on all 13 tested models in Q4_K_M
• llama.cpp needed a special build without repacking and managed 2.8 tok/s