Nemotron 3 Nano Omni vs MiniCPM-o 4.5: which open omni model?
NVIDIA's 31B omni model reads video, audio and images and answers in text. OpenBMB's 9B MiniCPM-o 4.5 also speaks. Compare memory, languages and license.
Verdict. Choose by what the model must say and how much hardware you have. Nemotron 3 Nano Omni has 31B parameters with about 3B active, a 256K context, video of up to 2 minutes and audio of up to an hour, and it answers in text only, in English only, under NVIDIA's Open Model Agreement, which allows commercial use. NVIDIA lists an 80 GB H100 for BF16, a 48 GB L40S for FP8 and a 32 GB RTX 5090 for NVFP4. MiniCPM-o 4.5 is a 9.37B dense model that also speaks, in English and Chinese, can take live video and audio while it talks, and uses Apache 2.0. OpenBMB's table lists 19.0 GB of GPU memory in BF16 and 11.0 GB in int4, but the model config sets only 40,960 positions and its full-duplex demo needs a 28 GB GPU. Both tables of scores are the developers' own and use different rivals, NVIDIA's older Nano VL V2 against OpenBMB's Gemini 2.5 Flash, GPT-4o and Qwen3-VL, so they cannot be compared. Test both on your own clips.
Specifications
| Specifications | Nemotron 3 Nano Omni | MiniCPM-o 4.5 |
|---|---|---|
| Developer | NVIDIA | OpenBMB |
| Released | April 28, 2026 | February 3, 2026 |
| Total parameters | 31B (Nano Omni 30B-A3B (BF16)) | 9.37B (9B (BF16)) |
| Active parameters | 3B | All |
| Architecture | Mixture of experts | Dense |
| Context window | 256K tokens | 40K tokens |
| Input | Text, Image, Audio, Video | Text, Image, Audio, Video |
| License | NVIDIA Open Model (Permissive) | Apache 2.0 (Permissive) |
| Memory for the smallest variant (4-bit) | 32 GB (Nano Omni 30B-A3B (NVFP4)) | 11 GB (9B AWQ (int4)) |
Performance claims come from the developer and have not been independently verified.
Pick Nemotron 3 Nano Omni if
- You have long inputs: NVIDIA lists a 256K context, audio of up to one hour and video of up to two minutes.
- You need text answers only, such as transcripts, summaries, JSON or tool calls, and English is enough.
- You have NVIDIA hardware with 32 GB or more, such as an RTX 5090 for NVFP4 or an L40S for FP8.
- You can run vLLM 0.20.0 on Linux and render PDFs to page images yourself.
Pick MiniCPM-o 4.5 if
- You need the model to talk back, in English or Chinese speech, with a configurable voice or a cloned one that you have consent for.
- You want live, full-duplex streaming, where the model hears and speaks at the same time and can speak up on its own.
- You have a 12 to 24 GB GPU or an Apple silicon Mac with 16 to 24 GB of memory.
- You want Apache 2.0 for both weights and code, and Chinese as well as English.