Musubi Labs open-sources PolicyLM-1.7B content moderation decision model

Musubi Labs releases PolicyLM-1.7B, a content moderation decision model for trust and safety teams, with weights released under Apache 2.0. It reads platform custom policies and messages, returns a 0-1 score for each category in a single forward pass, does not generate text, has a median latency of 35 ms on short chat messages on a 24 GB L4 and below 100 ms overall, supports multi-label and custom labels, and policy changes take effect immediately without retraining. The model is based on BidirLM-1.7B-Embedding, derived from Qwen3-1.7B-Base, and trained on public data such as the NVIDIA Nemotron Safety Guard Dataset v3, Alibaba XGuard-Train-Open-200K, and PolyGuardMix. It is evaluated on 19 languages, with English being the strongest and Tamil the weakest. The authors state it outperforms all comparison models under 20B parameters on their self-built policy benchmark, and a fine-tuned version is already running on a platform processing over one million messages per day.