A massive mixture-of-experts model that punches well above its active parameter weight — only 12B parameters are active at inference time despite the 120B total, making it surprisingly efficient to run. It handles long-context tasks with a 1M token window, which is unusually generous for an open-weight model. As a base model, it has no instruction tuning or chat formatting, so it requires fine-tuning or prompting expertise to get useful outputs.