More and more people have been using it every day recently OpenClaw、Hermes Agent this kind AI agenttools, from writing code and processing documents to automating just about anything, if you want to use a local model, earlier NVIDIA The newly released Nemotron 3 Nano Omni fully multimodal open-source model is definitely worth trying—it has only 30B total parameters, supports multimodal inputs including text, images, audio, and video, ranks first in 6 benchmarks, and is up to 9x faster than comparable open-source omni models. It is currently available for download on platforms like Hugging Face and Ollama.
NVIDIA Unveils Nemotron 3 Nano Omni: 30B Open-Source Multimodal MoE Model, Ranks #1 in 6 Major Benchmarks with Up to 9x Faster Efficiency Than Competitors
Nemotron 3 Nano Omni It is the flagship model in the Nemotron 3 series family for multimodal and agentic applications, featuring a 30B-A3B hybrid MoE architecture with a total of 300 billion parameters, but only 3 billion are activated per inference. Combined with Conv3D and EVS technologies, it supports a context length of up to 256K tokens. The biggest advantage of the MoE architecture is that it significantly reduces computational costs compared to similarly-sized dense models.
For its multimodal capabilities, Nemotron 3 Nano Omni supports text, images, audio, video, documents, charts, and tables simultaneously, and can even read GUI interface screenshots. The output is unified as text. This makes it particularly suitable for tasks such as document understanding, video summarization, and speech transcription. Additionally, it enables AI agents to operate computers by looking at the screen – the model can see the current GUI interface, identify elements on screen and state changes, and then decide what to click next.
NVIDIA also emphasized that this model is particularly useful for workflows involving mixed-media inputs, such as enterprise document analysis and compliance review.
In terms of performance testing, NVIDIA officially stated that Nemotron 3 Nano Omni achieved 6 first-place rankings across multiple multimodal benchmarks, with 5 explicitly named: MMlongbench-Doc and OCRBenchV2 (document understanding), WorldSense and DailyOmni (video understanding), and VoiceBench (speech understanding).
Compared with other open-source omni models, the gap becomes even more apparent. On NVIDIA B200 GPUs, Nemotron 3 Nano Omni achieves 9.2x system efficiency in video usage scenarios:
System efficiency in multi-document usage scenarios reaches 7.4x:
The figure below shows the multimodal accuracy of the previous generation Nemotron Nano VL V2 model and Nemotron 3 Nano Omni, with improvements across multiple benchmark tests:
Nemotron 3 Nano Omni is now available on Hugging Face, OpenRouter, build.nvidia.com, Ollama, and LM Studio, so if you can’t run it locally but want to try it,OpenRouter There is a free tier available, with $0 for both input and output, though there are rate limits (20 requests per minute, 200 requests per day). You can also set up a free API key through NVIDIA to use it (Click here for instructions）：
Source: KOCPC Chinese