NVIDIA's small multimodal reasoning model: 30B MoE, only 3B active, text+image+audio input. Free on OpenRouter. Designed for on-device, 4x faster than its predecessor, reasoning ON/OFF mode.
At a glance
Context
256K tokens
Parameters
30B MoE (3B active)
Input
text, image, audio, code
Specifications
- Status
- Released
- Date
- April 30, 2026
- Lab
- NVIDIA
- Origin
- US
- Type
- reasoning, multimodal
- Open weights
- Yes
- License
- NVIDIA Open Model License
- Context
- 256K tokens
- Parameters
- 30B MoE (3B active)
- Input
- text, image, audio, code
- Output
- text, code