Janas runs LLMs bigger than your RAM by streaming experts from SSD

Janas, launched in October 2026, lets users run MoE models that exceed available RAM by streaming experts from SSD on demand. A 70B MoE model ran on a 16GB RAM machine; the CLI is free and open source under MIT.

• Janas streams only the needed MoE experts from SSD, not the whole model
• A model normally needing ~100GB runs on a 16GB RAM machine
• Caching frequently used experts in RAM speeds up repeat runs
• CLI is free and MIT-licensed, supports open-weight MoE models only