NVIDIA Unveils Nemotron 3 Nano Omni

NVIDIA has unveiled Nemotron 3 Nano Omni. It's a multimodal LLM that processes video, audio, image, and text together, targeting enterprise Q&A, summarization, transcription, document intelligence, and GUI automation. It went up simultaneously on Build.Nvidia.com, Hugging Face, and NGC on April 28, 2026, and is available for commercial use under the NVIDIA Open Model Agreement. Qwen3-VL-30B-A3B-Instruct, Qwen3.5-122B-A10B, Qwen3.5-397B-A17B, Qwen2.5-VL-72B-Instruct, and gpt-oss-120b were used for improvement.

The key specifications are as follows.

• Architecture: Mamba2-Transformer hybrid MoE
• Composition: Nemotron 3 Nano LLM (30B A3B), CRADIO v4-H vision encoder, Parakeet speech encoder
• Parameters: 3.1×10^10 (31B A3B)
• Input: video (mp4, up to 2 minutes), audio (wav/mp3, up to 1 hour), image (RGB), text
• Video sampling: 1 FPS/128 frames guided for 1080p, 2 FPS/256 frames for 720p
• Context: up to 256k tokens
• Output: text, JSON, reasoning output, tool calling, word-level timestamps
• Language: English only

Deployment paths are also broad.

• Runtime: vLLM, NeMo, Megatron, NeMo-RL
• Inference runtime: TensorRT LLM, TensorRT Edge-LLM, llama.cpp, Ollama, SGLang
• Supported hardware: Ampere, Hopper, Blackwell, Lovelace families
• vLLM requires 0.20.0, and vllm[audio] must be installed to use audio.

NVIDIA also provided example run commands. For a typical 1-GPU environment, options like --max-model-len 131072, --reasoning-parser nemotron_v3, --enable-auto-tool-choice, and --tool-call-parser qwen3_coder are used, and --kv-cache-dtype fp8 is omitted for BF16. It notes that RTX Pro needs --moe-backend triton, and the NVFP4 + TP>1 combination needs --moe-backend flashinfer_cutlass.

Examples for DGX Spark (ARM64) and memory tuning are also included. In unified LPDDR5X memory environments, it guides lowering --gpu-memory-utilization and --max-model-len to reduce KV cache and memory pressure.