Musubi released PolicyLM-1.7B, an open-weight decision model designed to apply written moderation policies to content in under 50 milliseconds.
Musubi announced PolicyLM-1.7B on October 6, 2026, presenting it as a decision model for real-time content moderation. The company released the model with open weights, so organizations can run it themselves.
The model is designed to evaluate messages against moderation rules written in plain English and return a binary classification: whether content belongs in a specified category or not. Musubi says the process takes under 50 milliseconds. Unlike a conventional classifier that may need additional training to reflect policy changes, PolicyLM-1.7B is intended to apply revised rules without being retrained.
That approach could give platform teams a way to adjust moderation criteria as policies change, while labeling large volumes of content. Musubi says the model is meant to have costs and speed comparable to AI classifier systems already used for moderation on social platforms, while handling more complex rules without special training. Those are design claims reported with the announcement; the article did not provide benchmark results or independent performance testing.
Musubi co-founder and chief AI officer Filip Jankovic described the model as a way for product teams to get a more scalable and customizable picture of activity on their platforms. The company’s focus is proactive labeling, rather than having the model generate explanatory text or make moderation decisions beyond its configured categories.
Decision models have also drawn attention through TypeSafe AI’s Jev, released in September, and subsequent models from OpenAI and Amazon. These systems produce outcomes or probabilities rather than open-ended text. Musubi says its interest in the approach predates that recent attention, pointing to a 2024 project called GLiNER. PolicyLM-1.7B applies the decision-model format specifically to content moderation.