Musubi open-sources PolicyLM-1.7B under Apache 2.0 for policy-aware content moderation at classifier latency

What changed

Musubi released PolicyLM-1.7B on October 6 as a 1.7B decision model for real-time Trust & Safety moderation, with open weights under Apache 2.0. It reads a platform's own categories and rules alongside each message and returns a 0–1 score for every category in a single pass without generating an explanation, allowing policy edits without retraining a fixed classifier. Musubi reports a median of about 35 ms for short chat messages with six categories on a 24GB L4 and targets under 100 ms overall; it also reports leading results among the sub-20B models in its custom-policy benchmark. Those latency and accuracy figures are vendor-run tests. The model is text-only, has a 2,048-token context and was evaluated in 19 languages, with English strongest.