As decision-making models spread across industries, a company called Musubi came up with a new idea for making decision-making models work: content moderation. On Tuesday, Musubi announced a lightweight decision-making model built for real-time moderation called PolicyLM-1.7B, released in open weight.
The idea is to take a content policy written in plain English and apply it to messages within 50 milliseconds. Musubi’s models are designed to be cost and speed comparable to the AI classification systems that power moderation on most social platforms, but with the flexibility of modern LLMs to apply complex policies without special training. More importantly, the model requires no new training when the policy changes, allowing human policy setters to iterate as often as needed.
Musubi co-founder and chief AI officer Filip Yankovic believes this will give platform managers a way to proactively label content.
“Product teams want to better understand what’s happening on their platforms, especially as the amount of content grows exponentially,” Jankovic says. “Being able to label all of that in a very scalable and customizable way is very convenient.”
Decision-making models have been a hot topic in the AI world since TypeSafe AI’s Jev was released in September, quickly followed by competing decision-making models from OpenAI and Amazon. Instead of outputting text, the decision model outputs the probability of an outcome, but in this case, the model outputs a binary decision: whether the content is in the category or not. By restricting the model output to a predetermined set of choices, decision models can run faster and cheaper than large language models while preserving the flexibility of the transformer architecture.
One of the early use cases is reducing fraud with AI agents. Therefore, it is natural to apply the same technology to human fraud.
Notably, Yankovic says that interest in decision-making models predates Jev, dating back to a 2024 project called GLiNER (Generalist Model for Named Entity Recognition) that introduced much of the same technology.
Still, Musubi isn’t wary of the comparison. Rather, the company hopes to use the new interest in decision-making models to shine a light on 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 reads.
If you make a purchase through links in our articles, we may earn a small commission. This does not affect editorial independence.
