Change moderation rules without retraining: PolicyLM reads them on every call

Musubi released PolicyLM-1.7B with open weights to check messages against your rules.

PolicyLM checks messages without generating a response. When the rules change, the model reads a new instruction rather than being retrained.

For an application with comments, it serves as a separate moderation model that can run offline. Musubi published it on Oct 6, 2026. The model runs on CPU and Apple silicon, and the weights file takes up 3.5 GB. The Apache-2.0 license permits commercial use.

Rules as input. Calling `model.classify(message, policy)` returns scores from 0 to 1 for the specified categories. For example, an application can check a message against its own publishing rules. Instead of retraining after every edit, you pass the updated policy in the next call.

A custom policy can contain up to 16 categories and 1,662 tokens. The rules and message together must fit within 2,048 tokens. The `aegis` mode offers 23 built-in categories. The model receives the text of a single message, with no images or conversation history.

For custom policies, the available thresholds are `precision` at 0.335 and `balanced` at 0.275. A `cutoff` threshold can be set separately for each category. Musubi evaluated messages in 19 languages and tested the rules in English.

In Musubi's measurements on Oct 6, a short message with 6 categories took a median of 35 ms to process on an NVIDIA L4 with 24 GB of memory and 22 ms on an H100. In batch processing, the L4 handled 39.3 messages per second.

In a synthetic test of custom rules, Musubi compared PolicyLM with gpt-oss-safeguard-20B. On an H100, the smaller model responded faster but matched the correct answer less often. Accuracy was measured at a `cutoff` of 0.5.

| Model | Share of correct decisions | Latency on H100 | | --- | --- | --- | | PolicyLM-1.7B | 84.2% | 22 ms | | gpt-oss-safeguard-20B | 90.9% | 349 ms |

Running it locally. Musubi's official instructions from Oct 6 use the public `v1.2` release. A new virtualenv requires Python ≥3.10, transformers 4.57.6 and torch ≥2.5.1. Transformers 5.x is not yet supported.

```bash pip install "huggingface_hub>=0.34,<1.0" hf download musubilabs/policylm-1.7b --revision v1.2 --local-dir policylm-1.7b pip install -r policylm-1.7b/inference/requirements.txt python policylm-1.7b/inference/quickstart.py policylm-1.7b ```

The public `v1.2` release was also tested on a Mac. In the Hukuhuku review, Toolbox EP.57, on Oct 7, author 松尾亮 used 12 fictional messages in Japanese and 2 categories with rules in English. All 24 category decisions matched the predefined expectations.

Original source: [TechCrunch, Oct 6, 2026](https://techcrunch.com/2026/10/06/how-ai-decision-models-could-change-content-moderation).

PolicyLM returns category scores, and the application decides which messages to publish.