Janas brings SSD-streamed local AI to Windows
The experimental local AI engine now has Windows and Linux packages; its author says an 80B model runs on a 32 GB laptop by streaming experts from an SSD.
More model than memory
Janas has released Windows x86-64 binaries for its experimental local language-model engine, following its Linux package. Its October 12 change log says the Windows build uses the same source code, with platform-specific SSD reading and system services. The project streams mixture-of-experts weights from an NVMe drive as tokens need them, keeping frequently used experts in memory. That design lets the author run a 48 GB Qwen3-Next-80B model file on a laptop with 32 GB of RAM. A fresh Hacker News discussion on October 12 brought the project to wider attention; the repository itself dates to September, so this is a Windows packaging and current-interest story, not a new project launch.
Promising numbers, narrow evidence
The author reports 37 generated tokens per second for that model with draft-token assistance, and 24 without it, on one Intel Core Ultra 9 laptop with integrated graphics. The README compares its own measurements with llama.cpp on the same machine, but those numbers are not independent tests and do not predict results on other hardware. Janas explicitly calls itself an experiment: its interface may change, its stability has not been verified, and intensive disk and memory use can slow or disrupt a machine. The downloadable packages make the approach testable, while broader compatibility and sustained performance remain open questions.