Musubi announced Tuesday that it has released PolicyLM-1.7B, a lightweight decision model designed for real-time content moderation. The system operates with open weights and is built to process messages in under 50 milliseconds.
The approach
The model accepts content policies written in plain English and applies them to messages without requiring special training. This flexibility allows human policy-setters to update rules as needed without triggering new training cycles. The cost and speed of PolicyLM-1.7B match existing AI classifier systems used on most social platforms.
Filip Jankovic, co-founder and chief AI officer at Musubi, states the tool gives platform managers a way to label content proactively.
“Product teams just want a better understanding of what’s happening on their platform, especially as the amount of content is exponentially increasing,” Jankovic says. “Being able to label all of that in a very scalable, customizable way is extremely useful.”
Decision models have gained attention since TypeSafe AI released Jev in September, followed quickly by similar models from OpenAI and Amazon. Unlike large language models that output text, these systems return outcome probabilities. In this specific case, the model provides a binary judgement: the content either fits a category or it does not.
Limiting the output to predetermined choices allows the models to run faster and cheaper while keeping the transformer architecture. Musubi notes this technology is already used to control AI agent behaviour, making it a natural fit for managing human conduct.
Jankovic says his interest in these models predates Jev, originating from a 2024 project called GLiNER (Generalist Model for Named Entity Recognition) which deployed similar techniques. The company is not avoiding comparisons to Jev. Instead, it positions PolicyLM-1.7B as a version of that same model type, trained specifically for content moderation, which creators can run themselves.
