NVIDIA Nemotron Nano 2 VL

NVIDIA Nemotron Nano 2 VL is a 12B parameter open multimodal reasoning model developed by NVIDIA. It is designed for document intelligence, visual question answering, long video understanding, and complex multimodal reasoning. It builds on the Nemotron Nano V2 LLM and incorporates the RADIOv2.5 vision encoder with a hybrid Mamba-Transformer architecture.

Nemotron Nano 2 VL is optimized for:

• Multi-modal document intelligence (e.g., invoices, receipts, manuals)
• Multi-image reasoning (up to 5 document images)
• Visual question answering (VQA)
• Video understanding and summarization

The model supports a maximum of 128K tokens (input + output combined) at inference time.

Although training included sequences up to 311K tokens, the usable context length during inference is 128K tokens.

Yes. The model is released in:

• FP8
• FP4 (using Quantization-aware Distillation)
• BF16

• Performance degradation on text reasoning tasks after vision training (mitigated in later SFT stages)
• Repetition loops in reasoning-on mode for OOD tasks (addressed via reasoning budget control)
• Performance sensitivity in OCR tasks based on image preprocessing method

The model has 12.6 billion parameters.

No, fine-tuning is not currently supported on Fireworks AI.

It is available via on-demand deployment only, which provides dedicated GPUs with no rate limits.

Use of this model is governed by the NVIDIA Open Model License Agreement.