600 million-parameter real-time speech recognition + 4 billion-parameter multimodal safety guardrails—Nemotron 3.5’s combination punch lands perfectly. It can run on L40s, the 128k-context auditing is friendly, and enterprise customers will likely be more than happy to buy in.

CryptoWorld News reports that NVIDIA has open-sourced two specialized models: the 600 million parameter real-time multilingual speech recognition model Nemotron 3.5 ASR, and the 4 billion parameter multimodal safety guardrail model Nemotron 3.5 Content Safety, based on fine-tuning of Google's Gemma-3-4b-it. Nemotron 3.5 ASR supports adaptive punctuation and case prediction for over 40 languages, with streaming inference latency below 100 milliseconds, enabling high throughput when deployed on NVIDIA L40s GPUs. The content safety model supports 12 languages including Chinese and English, covering 23 safety categories, capable of evaluating user prompts, input images, and assistant responses within a 128k token context, and generating inference traces for enterprise auditing.