Musubi announced PolicyLM-1.7B on October 6, releasing a 1.7-billion-parameter content-moderation model with open weights under Apache 2.0. It is designed to evaluate messages against a platform’s own written policies, rather than restrict operators to a fixed set of moderation categories.

The model reads a policy and a message together, returning a score from zero to one for each category in a single pass. It generates no written explanation. Teams can change category rules without retraining, although Musubi says those instructions refine learned meanings rather than freely redefine them.

Fast decisions, with defined limits

Musubi reports a median latency of 35 milliseconds for short chat messages on a single 24 GB NVIDIA L4 GPU with up to six categories. That is a company-reported measurement under specific conditions, not a guarantee for every workload. The release targets live chat, direct messages, game lobbies and usernames. Its model card says the released model has not yet been tested on live traffic.

The model is text-only and handles one message at a time without conversation history. Its context window holds 2,048 tokens across the policy and message. Musubi evaluated messages in 19 languages, but says English performs best and all tested custom policies were written in English.

What developers get

The downloadable release includes inference helpers and two detection-threshold presets: precision, the default for settings where violations are uncommon, and balanced for catching more violations. Musubi recommends calibrating thresholds on a platform’s own content before deployment.

For developers, the concrete addition is a self-hostable policy-scoring component. Its lack of explanations and conversational context leaves a separate role for human review or larger models in appeals and nuanced enforcement decisions.