Musubi releases PolicyLM-1.7B, an open moderation model that scores a platform's own policy

Musubi has released PolicyLM-1.7B as open weights under Apache 2.0, targeting trust-and-safety teams that need a decision on short messages before a conversation moves on. The model reads a supplied policy together with the message and returns a score for each category, rather than writing a chatbot-style verdict.

In its launch explanation, Musubi positions the model for live chat, game lobbies, direct messages and usernames. The company says a custom fine-tuned version already runs on a platform handling more than a million messages per day. That production claim concerns a customized version and should not be treated as a measurement of the unmodified release on another community's content.

Policy labels can change without retraining

Teams describe their own categories in plain language, including rules for violations and exceptions. PolicyLM reads those instructions on each message and returns 0–1 scores in a single pass. A policy team can refine categories without waiting for another model-training cycle.

Musubi distinguishes that flexibility from complete freedom to redefine a label. The model retains meanings learned during training, with the supplied definitions layered on top. The publisher says instructions can shift scores but are not intended to redefine abuse as support. Teams should test the intended exceptions instead of assuming that a written rule guarantees the scoring behavior they want.

The release supports positive, pro-social categories as well as harmful-content labels. It can use custom rules or NVIDIA's Aegis taxonomy, which Musubi describes as a public set of 23 standard harm categories.

Latency claims include a specific test setup

Musubi reports a median of 35 milliseconds for short chat messages on a 24 GB L4 GPU with six categories, and describes latency below 100 milliseconds. It also says the model can run on a laptop or a single 24 GB GPU. The reported median depends on its stated message and category setup; it is not a promised response time for arbitrary text lengths or hardware.

The model generates no explanatory text. That removes the need to parse a written verdict, but also means moderators do not receive a rationale for an individual score. Musubi suggests larger models for appeals, bans, takedowns and unfamiliar judgment calls, with smaller specialized models handling high-volume labeling.

The company compares the decision-only approach with TypeSafe's Jev. Readers considering that distinction can examine the Jev walkthrough's content-moderation segment; it is a separate demonstration, not a benchmark comparison with PolicyLM.

Thresholds need calibration on real content

PolicyLM ships with precision and balanced cutoff presets. Musubi presents precision as the default for settings such as live chat where violations are rare. Balanced catches more, while a lower cutoff can support recall-first triage. Teams can set thresholds per category.

The launch advises calibrating on a sample of the team's own content before going live. That advice is important for communities whose language, context or rules differ from the publisher's tests. A score threshold defines what gets flagged, so deployment review needs to include both missed violations and harmless messages incorrectly caught.

The helper cleans some disguised text before scoring. Musubi says handling look-alike characters, spaced letters and base64 caught 10 to 12 points more disguised violations in its test set at the balanced cutoff, without more false flags. It warns that leetspeak often gets through. Franklin has not independently reproduced those tests.

The release has clear context and language limits

PolicyLM is text-only and evaluates one message at a time, without conversation history. Its context window is 2,048 tokens for policy and message together; the helper also scores long messages in windows. Benign messages that sound harmful, long text and many unrelated categories can increase false flags.

Musubi evaluated 19 languages, with English strongest and Tamil weakest. All custom policies in its tests were written in English. The quickstart currently requires transformers 4.57.6 rather than version 5.x. These constraints belong in an adoption checklist alongside speed and licensing, especially for multilingual teams or workflows that need conversational context.