Musubi Launches PolicyLM-1.7B for Real-Time AI Content Moderation

AI startup Musubi is applying the emerging class of decision models to content moderation with PolicyLM-1.7B, a small open-weight model designed to evaluate messages against policies written in plain English. The company announced the model Tuesday as a faster alternative to using large generative models for every moderation decision.

PolicyLM-1.7B reads a platform’s policy alongside each message and returns labels and scores indicating whether the content falls into defined categories. Musubi says policy teams can change or refine those categories without retraining the model, allowing moderation rules to be updated without waiting for a new classifier to be built.

A moderation model designed for speed

The 1.7-billion-parameter model is intended to operate at speeds closer to traditional moderation classifiers than large language models. Musubi reports a median latency of 35 milliseconds for short chat messages when testing six categories on a 24 GB Nvidia L4 GPU, with the broader system designed to stay under 100 milliseconds.

That speed matters for platforms handling large volumes of messages in real time. Instead of generating open-ended text, PolicyLM works within labels defined by the operator, allowing a service to classify content while retaining more flexibility than a conventional classifier trained around a fixed taxonomy.

Musubi co-founder and Chief AI Officer Filip Jankovic said product teams increasingly need scalable ways to understand what is happening across their platforms as content volumes grow. The model is also designed to classify positive or pro-social material, rather than being limited to identifying violations.

Decision models trade generation for bounded outputs

Decision models have attracted more attention since TypeSafe AI introduced Jev in September, followed by similar work from other AI companies. Instead of producing prose like a chatbot, these systems return structured outcomes such as choices, scores or probabilities that software can use directly.

PolicyLM applies that approach specifically to moderation. Limiting the output space can make the system smaller and faster while still allowing policies to be described in natural language rather than encoded entirely through rigid rules or a separately trained classifier.

Jankovic traces his interest in the approach further back than the recent wave of decision models. He points to GLiNER, a generalist model for named entity recognition, as an earlier example of using a compact bidirectional transformer to make flexible classifications without relying on large-scale text generation.

Open weights could make policy experimentation easier

Musubi released PolicyLM-1.7B with open weights under an Apache 2.0 license, allowing organisations to run the model themselves. The company says a custom fine-tuned version is already being used on a platform processing more than 1 million messages per day.

The approach gives moderation teams another option between fixed classifiers and much larger generative models. Rather than retraining a moderation system whenever a policy changes, operators can adjust the policy and labels supplied to the model, then apply the revised rules directly to incoming content.