Introducing PolicyLM-1.7B: a small, fast, open model that reads your content policy | Musubi

Musubi Labs released PolicyLM-1.7B, a 1.7 billion parameter model designed specifically for real-time content moderation. It reads your platform's actual policies and returns a 0–1 score for each category you define, all in under 100 milliseconds — fast enough for live chat, game lobbies, and direct messages where decisions need to happen before conversation moves on. The model is open-source (Apache 2.0), runs on a single 24 GB GPU or even a laptop, and handles over a million messages daily in production. Unlike fixed classifiers that require retraining whenever rules change, PolicyLM reads your policy with every message, so policy updates take effect instantly. Unlike large language models that generate text one token at a time and cost far more, PolicyLM makes a decision in one pass without any text generation. The core advantage is steerability without constant retraining. You define your own labels in plain language — as broad or specific as your community needs — and the model applies its learned understanding of harm to your taxonomy. It supports 19 languages (English strongest), offers two threshold presets ("precision" for rare violations, "balanced" for frequent ones), and can score all your traffic instead of just sampling it. The comparison table shows it sits between fixed classifiers (fast but rigid) and LLMs (flexible but slow). It's genuinely faster than both at its specific job: fixed classifiers run in tens of milliseconds but can't read policy changes, LLMs usually take hundreds of milliseconds or more, and PolicyLM hits under 100 milliseconds while actually understanding your rules. The model has real constraints worth knowing. It handles text only, one message at a time with no conversation history. Tamil was the weakest language tested. It can flag false positives on benign content that sounds harmful, and it doesn't explain its decisions — you just get scores. The team trained it on public datasets from NVIDIA and Alibaba plus synthetic data, starting from a Qwen3-based encoder. If your moderation needs are stable and map to standard taxonomies, a fixed classifier might be cheaper. If you're handling novel edge cases or appeals, a larger LLM makes sense. PolicyLM fills the gap for teams that need to make fast, policy-aware decisions on every message at scale.