A policy change in content moderation used to often mean retraining the model. Musubi’s PolicyLM-1.7B takes a different approach: write policies directly in English, apply them to messages, and return a binary judgment—whether something falls into a category—in 50 milliseconds. Its weights are open source.
Its speedup comes from narrowing the output: it doesn’t generate text, only indicates whether a category matches. By eliminating token-by-token generation, its speed and cost can approach those of traditional classifiers; policy changes don’t require retraining, so iteration on moderation rules no longer gets stuck at that step.
Its speedup comes from narrowing the output: it doesn’t generate text, only indicates whether a category matches. By eliminating token-by-token generation, its speed and cost can approach those of traditional classifiers; policy changes don’t require retraining, so iteration on moderation rules no longer gets stuck at that step.