Musubi's open PolicyLM-1.7B reads a platform's moderation rules at inference time

Musubi designed PolicyLM to sit between two kinds of moderation tool ^[7][8]. A conventional classifier is fast, but its categories are set during training. Redefining harassment, fraud or spam can mean new labeling and a retrain ^[7]. A general-purpose LLM can follow a written policy, but it is slower and costs more when it has to judge every message ^[8]. In my view PolicyLM takes the right approach for a filter that sees every post: it puts the policy in the input, so a team can edit a rule without retraining the model ^[2]. Up to six categories are scored in one pass, and Musubi reports a 35-millisecond median for a short message on an NVIDIA L4 ^[4].

A threshold turns each score into a decision ^[2]. The model card tells teams to calibrate that cutoff against labeled examples from their own platform ^[14]. The scores are computed against the policy text, so I'd expect a substantive rewrite of a rule to need its cutoff checked again. A score is also not a rationale ^[14], and a harassment score is hard to paste into an appeal reply.

Musubi's main comparison is against gpt-oss-safeguard-20B. The larger model's median on an H100 was 349 milliseconds ^[15], about 16 times PolicyLM's ^[21]. The same card's accuracy figures put PolicyLM's error rate at 15.8% against 9.1%. On Musubi's own test, it was wrong about 1.7 times as often ^[22].

For the custom-policy score to carry over, two things would have to be true. A platform's rules would have to resemble Musubi's synthetic ones, and its traffic would have to resemble sets the model met during development ^[16]. The public safety table has a separate caveat. Each competing model was run on its own taxonomy or on whichever setup scored higher, and PolicyLM's scores vary by test ^[17]. The latency figures are medians on short messages ^[4]. A moderation pipeline is sized for its slow tail and its long posts, and a platform would have to measure those itself.

The 19-language figure covers messages only. Every policy in Musubi's evaluations was written in English ^[19]. A platform that writes its rules in Spanish would be running outside the measured conditions. Musubi pitches the model as one component of a Trust & Safety toolkit, alongside moderation workflows and fraud detection ^[13].

The founders say they started Musubi after running into the cost of malicious users and the limits of the tools platform teams had ^[9]. CEO Tom Quisel spent a decade building Trust & Safety systems and was CTO at Grindr and OkCupid ^[5]. Filip Jankovic, the chief AI officer, led data science at Evidation Health and traced the approach to a 2024 GLiNER project ^[6]. He told TechCrunch that as the amount of content rises, product teams need a clearer view of what goes on across their platforms ^[12]. Musubi says its customers include Bluesky, Stocktwits and Bumble ^[11].