Musubi has introduced a new lightweight decision model for real-time content moderation named PolicyLM-1.7B. This model applies a content policy written in plain English to messages in less than 50 milliseconds. Notably, it operates at a cost and speed similar to the AI classifier systems commonly used on social platforms. Its primary advantage is its flexibility; it can manage complex policies without requiring special training and does not need retraining when policies change. This allows human policy-makers to make adjustments whenever necessary.
Filip Jankovic, co-founder and chief AI officer at Musubi, notes that this feature enables platform managers to label content proactively. Jankovic explains, “Product teams just want a better understanding of what’s happening on their platform, especially as the amount of content is exponentially increasing. Being able to label all of that in a very scalable, customizable way is extremely useful.”
Recently, decision models have gained attention, particularly following the launch of Typesafe AI’s Jev in September. This model generated interest in decision models, which, unlike traditional systems that create text, yield outcome probabilities. In the realm of content moderation, the decision model provides a binary assessment: either the content fits a certain category or it does not.
This binary classification approach lets decision models run faster and at a lower cost than large language models while retaining the flexibility of the transformer architecture. One early use case has been reining in misbehavior by AI agents, making content moderation a natural application of the same technology.
Interestingly, Jankovic’s interest in decision models goes back further than Jev, to a 2024 project named GLiNER that employed similar techniques. Musubi is eager to use that momentum to draw attention to content moderation specifically. The company claims, “If Jev caught your eye, PolicyLM-1.7B is the same kind of model, trained specifically for content moderation, that you can run yourself.”
The introduction of PolicyLM-1.7B highlights a shift in content moderation, where the ability to swiftly adapt to changing policies empowers platform managers to label content with unprecedented efficiency. This flexibility not only addresses the growing volume of user-generated content but also positions decision models as a viable alternative to traditional systems, offering a blend of speed, cost-effectiveness, and adaptability that could redefine how platforms approach moderation.
