Musubi drops PolicyLM-1.7B, an open-weight mod bot that reads your rules in 35ms

PolicyLM-1.7B is an open-weight 1.7B classifier built for teams that need instant content moderation decisions without the endless cycle of dataset relabeling and retraining. Instead of relying on a hardcoded taxonomy, it reads your custom plain-text rules alongside every incoming message and enforces policy updates on the fly.

Traditional classifiers are fast but rigid, meaning any policy tweak forces you to annotate fresh data and retrain. Large language models can understand nuanced guidelines, but generating token-by-token text is far too slow and expensive for high-volume live chat. PolicyLM-1.7B takes a decision-first approach: it runs a single forward pass and outputs a 0 to 1 confidence score across all your categories without generating a single word of text.

What it can do

Musubi built PolicyLM-1.7B for real-time environments like gaming lobbies, direct messages, live streams, and input guardrails sitting directly in front of LLM applications.

It is lightweight enough to run on a single 24GB GPU (like an Nvidia L4) or a modern laptop. In internal benchmarks handling short chat messages across six categories on an L4, it clocks a median latency of 35ms and stays well under 100ms overall. That makes 100% traffic inspection practical without resorting to sampling.

You can leverage Nvidia's Aegis taxonomy of 23 harm categories out of the box or feed it custom instructions in plain language, such as allowing game-related trash talk while banning player harassment. It ships with presets for high precision (ideal for clean chat feeds) and balanced recall.

Under the hood, the model is derived from BidirLM-1.7B-Embedding and Qwen3-1.7B-Base. On custom-policy benchmarks, it achieved around 83% accuracy, outperforming every tested model under 20B parameters.

How to run it and known limits

The model weights are available on Hugging Face under the Apache 2.0 license. The official quickstart relies on transformers 4.57.6, and version 5.x is not supported yet. The inference helper includes handy pre-processing to peel back evasive tricks like look-alike characters and Base64 padding.

Keep in mind its architectural limits. The context window is capped at 2048 tokens for your policy and message combined. It evaluates single messages only, meaning it has no conversational memory, and it outputs raw scores rather than written explanations. Testing covered 19 languages, with English being the most reliable and Tamil the weakest.

Why it matters

If you run an interactive chat platform or need a snappy guardrail to protect an AI assistant from toxic prompts, PolicyLM-1.7B deserves a look. It offers a practical middle ground: the sub-100ms throughput of a legacy classifier paired with the policy flexibility of a modern LLM.

We think this non-generative, decision-only architecture is where high-throughput safety tooling is headed. It will not replace giant frontier models for handling user ban appeals, but for triage at the front door, it cuts both cloud costs and engineering turnaround dramatically.