Musubi has introduced PolicyLM-1.7B, a lightweight decision model designed for real-time content moderation. The model is released with open weights and aims to apply content policies written in plain English to messages in under 50 milliseconds.
The system is intended to match the speed and cost efficiency of existing AI classifiers while offering the flexibility of a modern large language model. It allows for policy updates without requiring additional training, enabling human teams to iterate on rules as needed.
Unlike standard generative models, this decision model outputs binary judgments regarding whether content falls into a specific category. This architecture allows the tool to operate faster and at a lower cost than traditional large language models.
Musubi co-founder Filip Jankovic noted that the company's work on decision models builds upon techniques used in a 2024 project called GLiNER. The firm positions PolicyLM-1.7B as a self-hosted alternative for platforms seeking scalable content labeling.
