Musubi releases a moderation model that reads each platform’s rules
Musubi released PolicyLM weights for content moderation, pairing each message with the platform’s written rules. Teams can change those rules without another training cycle and run the model on their own infrastructure. It assigns category scores rather than generating explanations. Deployment requires thresholds calibrated on local content, and the model evaluates individual text messages without conversation history.
Artificial Intelligence··Midday
The model applies the rules supplied with each message
Musubi, a developer of content-moderation AI, released PolicyLM-1.7B on 6 October with weights under Apache 2.0. The model receives a platform’s written rules alongside a message and classifies the content. Teams can define categories in natural language and revise them without retraining. Intended uses include live chats, game lobbies, direct messages and usernames on online platforms.[1], [2]
The output is a score from zero to one for each category, rather than a written explanation or a conversational response. Several labels can be evaluated in one pass. The platform sets thresholds for using those scores in decisions about content. Downloadable weights let teams run and fine-tune the model themselves; Musubi also offers managed deployment.[1]
Deployment thresholds depend on the platform’s content
Musubi recommends calibration using a sample from the platform’s own content before deployment. One default preset prioritizes precision for live conversations where rule violations are uncommon. A balanced preset is intended for review queues with more frequent violations or greater consequences from misses. Teams wanting broader triage can lower the cutoff to send more potentially problematic material for attention.[1]
Policy and message share a context window of 2,048 tokens, the units processed by the model. PolicyLM handles text and assesses one message at a time, without conversation history. Musubi identifies benign wording that sounds harmful, long inputs and many unrelated categories as potential sources of false flags. Its helper normalizes look-alike characters and spaced letters and splits long messages into windows.[1]
Local weights come with language and training limits
Musubi says it evaluated PolicyLM in 19 languages, with English strongest and Tamil weakest. The custom policies in those tests were all written in English. Co-founder and chief AI officer Filip Jankovic describes the product as a way for platform teams to customize moderation as message volumes grow. The language evaluation is Musubi’s own testing. Human policy-setters determine the categories, and the model reads those rules during use.[1], [2]
The model starts from BidirLM-1.7B-Embedding, an encoder derived from Qwen3-1.7B-Base. Training combines safety datasets from NVIDIA and Alibaba, PolyGuardMix and generated examples. Weights and the model card are hosted on Hugging Face. Musubi has also joined the ROOST Model Community, a safety-model collaboration, and published a joint guide on choosing and routing moderation models for different applications.[1]
