Nemotron 3.5 turns content safety from a filter into a policy engine

NVIDIA has released Nemotron 3.5 Content Safety on Hugging Face. The announcement describes a 4B model built on Google Gemma 3 4B IT, with a 128K context window, multimodal input, custom policy enforcement and an optional THINK mode for reasoning traces.

One safety output now covers the prompt, image and response

Nemotron 3.5 takes a user prompt, an optional image and an optional assistant response as one context. NVIDIA is targeting cases where the violation emerges from the combination of modalities.

The model follows Nemotron 3 Content Safety from March 2026. Version 3.5 adds policy specification at inference time and can output reasoning before the final safe or unsafe verdict.

Enterprise safety starts where universal taxonomies stop working

The important part is that a company can send its own policy with the input. A healthcare app, financial chatbot, IDE and children’s product do not share one risk model.

NVIDIA lists 12 languages with explicit training coverage and roughly 140-language zero-shot generalization inherited from the Gemma 3 base model. That helps global deployment, but it requires evals in specific markets.

A reasoning trace is not an audit until someone can trust it

A reasoning trace in a safety model sounds like auditability, but it is still model output. It can help a reviewer, but it does not solve calibration, bias or false negatives by itself.

Custom policy enforcement also does not mean governance is finished. Someone still has to write, version, test and resolve conflicts between local rules and the global taxonomy.

Evals outside the comfortable taxonomy will decide

The next things to watch are the model card, dataset, benchmarks against the Aegis 2.0 taxonomy and deployments with real custom policies. The blog also mentions a safety dataset, which helps reproducibility.

The real test will come in less comfortable languages, multimodal edge cases and regulated verticals.

Lilith's verdict

Nemotron 3.5 is not just a guard at the door saying yes or no. It is an attempt to hand that guard a company manual and a camera, but someone still has to check whether he read the right page.