PolicyLM-1.7B

PolicyLM-1.7B is an open-weights moderation classifier from Musubi, released October 2, 2026 under Apache 2.0. It is a fine-tune of BidirLM-1.7B-Embedding, trained on NVIDIA's Nemotron-Safety-Guard-Dataset-v3, Alibaba's XGuard-Train-Open-200K, and ToxicityPrompts/PolyGuardMix. Given a message and a policy, it returns a 0-1 score for every category of the policy in a single pass, up to 16 categories. Two modes are supported: explicit mode, where you write your own categories as short rules at inference time with no retraining, and aegis mode, which uses the 23 built-in categories from NVIDIA's Aegis 2.0 taxonomy for generic safety screening.

The design target is live decisioning: Musubi reports a 35ms median per short chat message on a single 24GB NVIDIA L4 with six categories (22ms on an H100), and the model also runs on laptop CPUs and Apple Silicon. It is multilingual (evaluated across 19 languages), self-contained with no trust_remote_code requirement, and works offline. Musubi's own benchmark claims it beats every sub-20B model they tested on custom-policy tasks. Out of scope by design: it gives scores, not written reasons, so it is not a substitute for appeals or takedown workflows, and it is explicitly not a child-safety enforcement tool.