Nvidia ships Nemotron 3 Nano Omni: free model, paid hardware

nvidia keeps dropping free models. they are just ads for their hardware

nvidia shipped nemotron 3 nano omni opensource today. there's one thing worth flagging

the target user isn't a solo builder. it's enterprises: contract analysis, delivery verification, document processing, gui agents for it teams. and the reason is straightforward – enterprises are the ones who can't send confidential contracts, customer footage, and delivery addresses to openai for legal reasons. they need on-premise. and they're also the ones who buy racks of nvidia gpus to run it on

that's the through-line. nvidia open-sources a capable model, enterprises adopt it for compliance reasons, enterprises buy the hardware to run it properly. palantir, foxconn, docusign, oracle are already on board

the full omni experience – proprietary nvfp4 quantization, audio encoder, vision encoder – that's nvidia hardware only. you can try it for free, but to run it properly at scale you need their stack. which is exactly the point

the model itself:
- 30b parameters
- 21gb in quantized form
- 256k context window

handles video up to 2 minutes, audio up to an hour, images, and text – all in one call. benchmark numbers are solid: 72 on video understanding, 89 on voice tasks, and computer-use scoring nearly 50% which is actually usable for real workflows

we already tested it – transcription works well