According to Dongcha Beating monitoring, NVIDIA has open-sourced two dedicated models for intelligent agent ecosystem optimization. These are the Nemotron 3.5 ASR real-time multilingual speech recognition model with 600 million parameters, and the Nemotron 3.5 Content Safety multimodal secure fence model with 4 billion parameters fine-tuned based on Google's Gemma-3-4B-it.
The Nemotron 3.5 ASR speech recognition model adopts a Cache-aware architecture that supports streaming-awareness, enabling adaptive punctuation and case prediction in over 40 languages. It achieves a single-step streaming inference latency of less than 100 milliseconds. In performance tests by the third-party Pipecat team, deployment on NVIDIA L40S GPUs demonstrated high throughput. Developers can balance latency and accuracy by dynamically configuring the Chunk Size.
The Content Safety review model Nemotron 3.5 Content Safety natively supports 12 languages including Chinese and English, covering 23 safety categories. It can simultaneously evaluate user prompts, input images, and assistant replies within a single 128K token context window. Apart from determining policy violations, the model can also generate Reasoning Traces along with output classification results, facilitating enterprise audits of the security decision-making logic.
