Musubi launches a decision model built for content moderation

Musubi on Tuesday announced PolicyLM-1.7B, a lightweight decision model built specifically for real-time content moderation, released with open weights. The model takes a content policy written in plain English and applies it to messages in under 50 milliseconds — matching the cost and speed of the AI classifier systems that already power moderation on most social platforms. Because it carries the flexibility of a modern large language model, though, it can apply complex policies without needing special training for each one. Just as important, it doesn’t need retraining when a policy changes, so the humans setting moderation rules can iterate as often as they need.

Musubi co-founder and chief AI officer Filip Jankovic frames the model as a way for platform managers to label content proactively rather than reactively. “Product teams just want a better understanding of what’s happening on their platform, especially as the amount of content is exponentially increasing,” he says — arguing that being able to label that volume of content in a scalable, customizable way is what makes the approach valuable.

Decision models have drawn attention across the AI industry since TypeSafe AI released its model, Jev, in September, quickly followed by competing decision models from OpenAI and Amazon. Rather than generating text, a decision model outputs outcome probabilities; PolicyLM-1.7B specifically returns a binary judgment — content either falls into a given category or it doesn’t. Restricting output to a fixed set of choices lets these models run faster and cheaper than general-purpose LLMs while keeping the flexibility of the underlying transformer architecture. One early use case for the category has been reining in misbehavior by AI agents, which Musubi is now extending to policing human behavior instead.

Jankovic says his own interest in decision models predates Jev, tracing back to a 2024 project called GLiNER (Generalist Model for Named Entity Recognition) that used many of the same techniques. Musubi isn’t shying away from the comparison to Jev — if anything, the company is using the attention around decision models to spotlight content moderation specifically. Its product announcement pitches PolicyLM-1.7B as the same kind of model as Jev, just trained for moderation and available to run yourself.

Why it matters

If moderation policy can change without retraining the underlying model, platforms can adjust rules as fast as regulation or public pressure demands — turning decision models from an AI-agent-governance tool into infrastructure for human content moderation too.