NVIDIA: Nvidia Nemotron Nano 12B V2 VL

NVIDIA's Nemotron Nano 12B 2 VL is a multimodal vision-language model built on a hybrid Mamba-Transformer architecture, combining the efficiency of state-space models with the flexibility of attention-based processing. This 12.6-billion-parameter model draws from the Nemotron-Nano-V2 large language model foundation and incorporates NVIDIA's C-RADIOv2-H vision encoder to weave visual and textual understanding together. Designed for document intelligence, video comprehension, multi-image analysis, and visual question answering, it handles extensive multimodal sequences without losing coherence, making it particularly effective for tasks like OCR, chart reasoning, and parsing complex visual documents.

The model is available as a BF16 checkpoint on Hugging Face, giving researchers and developers direct access to experiment with fine-tuning and deployment. It is optimized for high-throughput inference on NVIDIA hardware and performs efficiently even in single datacenter GPU setups. On benchmarks like OCRBench v2, it achieves leading results with throughput up to 2.5 times higher than comparable approaches, which speaks to its practical advantage in real-world workloads. Teams apply it to content analysis pipelines, document processing workflows, and anywhere that demands fast, accurate multimodal reasoning at scale.