ai/nemotron3 repository overview

⁠NVIDIA Nemotron 3 Nano Omni 30B A3B Reasoning

NVIDIA Nemotron 3 Nano Omni is a multimodal large language model that unifies video, audio, image, and text understanding to support enterprise-grade applications. It extends the Nemotron Nano family with integrated video and speech comprehension, Graphical User Interface (GUI) automation, Optical Character Recognition (OCR), and speech transcription capabilities. The model enables end-to-end processing of rich enterprise content such as meeting recordings, media assets, training videos, and complex business documents.

Built on a Mamba2-Transformer Hybrid Mixture of Experts architecture, Nemotron 3 Nano Omni delivers powerful multimodal understanding while maintaining efficiency through its active parameter design. The model supports reasoning mode with chain-of-thought capabilities, tool calling, JSON output formatting, and handles context lengths up to 256k tokens. It was developed using a comprehensive training pipeline involving over 354 million data points across text, audio, image, and video modalities.

This model is available for commercial use under the NVIDIA Open Model Agreement and can be deployed across various platforms including NVIDIA Ampere, Hopper, Blackwell, and edge devices like Jetson Thor.

⁠Characteristics

⁠Using this model with Docker Model Runner

docker model run nvidia-nemotron-3-nano-omni-30b-a3b-reasoning

For more information, check out the Docker Model Runner docs⁠.

⁠Benchmarks

Detailed benchmark results are available through NVIDIA's official documentation. The model has been evaluated across multiple modalities:

The model supports various quantization formats (BF16, FP8, NVFP4) and GGUF formats for efficient deployment across different hardware configurations.

⁠Links

• Unsloth GGUF Repository⁠
• NVIDIA Official Repository⁠
• NVIDIA Open Model Agreement⁠
• Build.NVIDIA.com⁠
• NGC Catalog⁠
• Jetson AI Lab⁠

⁠Considerations

• English-only support: While the model has multilingual capabilities for some languages (French, Spanish, Italian, German, Japanese, Chinese), primary optimization is for English content
• Reasoning mode: The model defaults to reasoning mode with chain-of-thought output; disable via enable_thinking: false for direct responses
• Video processing: Recommended frame sampling is 2 FPS with 128-512 frames depending on GPU memory (80GB+: 128-512 frames, ≤40GB: 64-256 frames)
• Audio format: Audio inputs should be resampled to 16 kHz; supports WAV and MP3 formats up to 1 hour
• PDF handling: The model processes images, not raw PDF files; pages must be rendered to PNG/JPEG before submission
• Inference frameworks: Supports vLLM 0.20.0+, TensorRT-LLM, TensorRT Edge-LLM, SGLang, llama.cpp, and Ollama
• Hardware requirements: Optimized for NVIDIA GPUs including Ampere (A100), Hopper (H100/H200), Blackwell (B200), Lovelace (L40S), and edge devices (Jetson Thor, DGX Spark)
• Quantization options: Available in BF16 (~62GB), FP8, NVFP4, and various GGUF quantization levels for different deployment scenarios
• Memory management: Unified memory architectures (like DGX Spark) may require tuning gpu-memory-utilization and max-model-len parameters
• Tool calling: Supports XML-based function calling format compatible with qwen3_coder parser
• Safety filtering: Model includes built-in content safety measures; additional filtering may be required for production deployments

⁠Generated by

This model card was automatically generated using cagent-action⁠. Want to learn more about Docker Model Runner? Check out the project repository: https://github.com/docker/model-runner⁠.