Nvidia has released Nemotron 3.5 with a customizable multimodal safety framework designed specifically for enterprises running models on local infrastructure. Unlike cloud-based moderation APIs that require data transmission to third-party services, Nemotron's safety layer runs entirely on-device, enabling organizations to enforce proprietary or region-specific content policies. The framework operates as a composable safety classifier system that evaluates both text and image inputs against configurable policy parameters, allowing financial institutions, healthcare providers, and government agencies to maintain data sovereignty while deploying powerful language models internally. This addresses a critical gap for organizations operating under strict data residency requirements—particularly those subject to GDPR in Europe or equivalent regulations in other jurisdictions where sending user interactions to external vendors creates compliance friction.
The safety framework was validated against EVA-Bench 2.0, an expanded evaluation suite covering 213 scenarios across three domains with 121 tools, and uses a hybrid approach combining learned classifiers with retrieval-augmented filtering mechanisms. Rather than relying on fixed blocklists, Nemotron's safety controls allow administrators to define thresholds and behaviors per use case; a customer service deployment might tolerate different language patterns than a financial advisory system. The system integrates directly with Ollama and llama.cpp workflows, meaning organizations already running local LLMs can layer safety controls without architectural changes. Nvidia has positioned this as an alternative to Together AI's managed safety modules and Replicate's moderation APIs, with the key distinction that enterprises retain complete control over policy logic and can audit decision-making without third-party involvement.
The framework is available now as part of Nvidia's open-source release, with no licensing restrictions preventing local deployment. Enterprise customers including firms in regulated industries have already begun evaluating the system for production use, where the ability to customize safety policies for domain-specific contexts—such as medical terminology in healthcare applications or financial jargon in banking—has proven essential. By enabling safety as a pluggable, locally-controllable layer rather than a black-box service dependency, Nemotron 3.5 shifts the control dynamic back to organizations deploying their own infrastructure, making self-hosted LLMs more viable for sectors where external API dependencies create unacceptable compliance or operational risk.