NVIDIA Releases Nemotron 3 Nano Omni: 30B-A3B Multimodal Model With 100+ Page Document Support
NVIDIA released Nemotron 3 Nano Omni, a 30B-A3B Mixture-of-Experts model that processes text, images, video, and audio. The model uses a hybrid Mamba-Transformer architecture with 128 experts and achieves 65.8 on OCRBenchV2-En and 72.2 on Video-MME, while delivering up to 9x higher throughput on multimodal tasks compared to alternatives.
Nemotron-3-Nano-Omni-30B-A3B — Quick Specs
NVIDIA released Nemotron 3 Nano Omni on April 28, 2026, a multimodal model that processes text, images, video, and audio in a unified architecture. The model uses a 30B-A3B Mixture-of-Experts backbone with 128 experts and top-6 routing.
Architecture and Scale
Nemotron 3 Nano Omni combines three encoder systems: C-RADIOv4-H for vision, Parakeet-TDT-0.6B-v2 for audio, and the Nemotron 3 Nano 30B-A3B language model. The architecture interleaves 23 Mamba selective state-space layers, 23 MoE layers with 128 experts, and 6 grouped-query attention layers.
For vision processing, the model supports dynamic resolution from 512x512 (1,024 patches) to 1840x1840 (13,312 patches) at native aspect ratio. Video processing uses Conv3D tubelet embedding that fuses consecutive frame pairs, halving the number of vision tokens.
Benchmark Performance
According to NVIDIA, Nemotron 3 Nano Omni achieves:
- 65.8 on OCRBenchV2-En (versus 61.2 for its predecessor Nemotron Nano V2 VL)
- 57.5 on MMLongBench-Doc
- 72.2 on Video-MME
- 89.4 on VoiceBench
- 5.95 word error rate on HF Open ASR
- 57.8 on ScreenSpot-Pro for GUI understanding
- 47.4 on OSWorld for computer use tasks
The model leads Qwen3-Omni 30B-A3B on most benchmarks, including document understanding (57.5 vs 49.5 on MMLongBench-Doc) and voice interaction (89.4 vs 88.8 on VoiceBench).
Throughput Claims
NVIDIA claims Nemotron 3 Nano Omni delivers up to 9x higher throughput and 2.9x faster single-stream reasoning speed on multimodal use cases compared to unspecified alternatives. The company states 7.4x higher system efficiency for multi-document workloads and 9.2x for video use cases compared to "other open omni models with the same interactivity."
Training Approach
The training recipe uses staged multimodal alignment and context extension, followed by preference optimization and multimodal reinforcement learning. The model can process 100+ page documents and includes an Efficient Video Sampling (EVS) feature that drops redundant video tokens after the vision encoder to reduce latency.
Availability
Nemotron 3 Nano Omni is available on Hugging Face in BF16, FP8, and NVFP4 formats. Pricing information was not disclosed.
What This Means
Nemotron 3 Nano Omni represents NVIDIA's entry into the competitive omni-modal space, positioning against models like Qwen3-Omni with a focus on enterprise document processing and computer use tasks. The hybrid Mamba-Transformer-MoE architecture is a notable architectural choice that differs from pure attention-based approaches, though real-world deployment efficiency will depend on framework support for these specialized layers. The strong document understanding scores (65.8 on OCRBenchV2) and computer use capabilities (47.4 on OSWorld) suggest practical applicability for enterprise workflows, though independent verification of throughput claims and production performance remains needed.
Related Articles
Moonshot AI Open-Sources Kimi K3 Weights After Model Matched GPT-5.6 Sol on Benchmarks
Moonshot AI has released open weights, a technical report, and supporting infrastructure for Kimi K3, a model that claims 2.5x more intelligence per unit of compute. Independent testing found notable gaps in cybersecurity and math performance compared to Western frontier models.
Microsoft Releases Fara1.5-27B, a 27B Vision-Only Web Browsing Agent with 262K Context
Microsoft Research AI Frontiers has released Fara1.5-27B, a 27-billion-parameter multimodal agent that completes web tasks by reading screenshots and emitting click/type/scroll commands. The model, fine-tuned from Qwen3.5-27B, ships under MIT license with a 262K-token context window and is designed to run alongside Microsoft's MagenticLite sandbox.
Microsoft Launches MAI-Cyber-1-Flash Security Model, Still Routes Hard Cases to OpenAI's GPT-5.4
Microsoft has released MAI-Cyber-1-Flash, a compact cybersecurity model built into its MDASH multi-agent system that scores 96 percent on the CyberGym benchmark. The setup handles 90 percent of security tasks in-house but still hands off difficult cases to OpenAI's GPT-5.4.
Microsoft Launches MAI-Cyber-1-Flash, Its First Cybersecurity Model, With Agentic Security Platform Perception
Microsoft has launched MAI-Cyber-1-Flash, its first cybersecurity-specialized model, alongside Perception, an agentic platform for automated threat detection and remediation. The company claims the model outperforms rivals from Anthropic, Google and OpenAI on the Cyber Gym benchmark, though no independent scores have been published.
Comments
Loading...