Decision models are gaining traction across the industry, and one company, Musubi, has a fresh take on how to apply them: content moderation. On Tuesday, Musubi introduced a lightweight decision model built for real-time moderation, dubbed PolicyLM-1.7B, which comes with open weights. The concept is straightforward—take a content policy written in everyday English and enforce it on messages in under 50 milliseconds. Musubi's model is engineered to match the cost and speed of the AI classifier systems that handle moderation on most social platforms, but thanks to the adaptability of a modern LLM, it can enforce intricate policies without requiring specialized training. Perhaps more crucially, the model doesn't need retraining when policies shift, giving human policy-setters the freedom to iterate as often as necessary. According to Musubi co-founder and chief AI officer Filip Jankovic, this provides platform managers with 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 drawn significant attention in the AI space ever since Typesafe AI released Jev in September, with competing decision models from OpenAI and Amazon arriving shortly after. Rather than generating text, a decision model produces outcome probabilities—in this instance, a binary verdict: either the content falls into the category or it doesn't. By constraining the model's output to a predetermined set of choices, decision models can operate faster and more cheaply than large language models, all while preserving the flexibility of the transformer architecture. One early application involves curbing misbehavior by AI agents, so extending the same approach to human misbehavior is a natural next step. Notably, Jankovic says his fascination with decision models goes back further than Jev, tracing it to a 2024 project called GLiNER (Generalist Model for Named Entity Recognition) that employed many of the same techniques. Even so, Musubi doesn't shy away from the comparison. In fact, the company is keen to leverage the renewed interest in decision models to draw attention to content moderation. "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 product announcement states.