Musubi’s PolicyLM-1.7B applies plain-English moderation rules in under 50 milliseconds

TechCrunch reports that Musubi has released PolicyLM-1.7B, an open-weight decision model built for real-time content moderation. Users write a content policy in plain English, and the model applies it to messages in under 50 milliseconds, at a cost and speed comparable to the classifiers most social platforms already use. Traditional classifiers need retraining when rules change; Musubi says its model does not, so policy writers can revise as often as needed. Instead of generating text, it returns a binary call on whether content falls into a category.

Decision models drew wide attention after TypeSafe AI released Jev in September, followed by competing models from OpenAI and Amazon (Amazon’s clone). Musubi co-founder and chief AI officer Filip Jankovic says his interest predates Jev, going back to a 2024 named-entity recognition project called GLiNER that used many of the same techniques. Other open decision models include Liquid AI’s d1 series (Liquid AI d1) and Nokia’s AnyJev (AnyJev).

Moderation has been stuck between cheap, rigid classifiers that miss nuance and expensive human review that cannot keep pace. A model that applies a written policy quickly without retraining could let smaller platforms run thoughtful policies instead of copying large platforms’ blunt filters. Open weights let communities and startups run and inspect it themselves. If the policy is a plain-English document, it can be read, debated and published, which beats a black-box classifier whose rules live only in training data.

Jankovic frames the value as letting product teams understand what happens on their platforms and label everything as content volume grows. That is useful, and it also describes comprehensive automated labeling of every message on a service. Cheap, instant, rewritable policies make it easy to change what gets flagged overnight, quietly and at scale; the same tool that refines harassment rules could target a topic, political view or community with a few sentences.

A binary yes-or-no also hides how close a call was. Whether platforms keep human appeal paths for borderline decisions, publish the policies they feed the model and log changes will determine whether the approach makes moderation fairer or merely faster. Independent accuracy and bias evaluations on real moderation data would help. More in our AI coverage, tools coverage and regulation coverage.