Nemotron 3.5 Lightning (free)

NVIDIA-Nemotron-Nano-9B-v2-free is a large language model trained from scratch by NVIDIA. Designed as a unified model, it efficiently handles both reasoning and non-reasoning tasks to respond to a wide range of user queries. With a generous context length of 128,000 tokens, it is highly capable of processing long and complex documents.

Developed by Nvidia, Nemotron-Nano-12B-V2-VL-Free is a 12-billion-parameter open multimodal reasoning model designed for video understanding and document intelligence. It features a powerful 128,000-token context length to handle extensive visual and textual inputs. The model utilizes a hybrid Transformer-Mamba architecture, successfully combining transformer-level accuracy with Mamba's structural advantages.

NVIDIA Nemotron 3 Super is a 120B-parameter open hybrid MoE model built on a hybrid Mamba-Transformer architecture. Activating just 12B parameters, it delivers maximum compute efficiency and accuracy for complex multi-agent applications. Additionally, it features an expansive context length of 262,144 tokens to handle large-scale inputs seamlessly.

Developed by Nvidia, NVIDIA Nemotron™ 3 Nano Omni is a 30B-A3B open multimodal model designed to function as a perception and context sub-agent in enterprise agent systems. Supporting a massive context length of 256,000 tokens, it accepts and processes inputs across text, images, and video. This model provides powerful multimodal comprehension tailored for complex enterprise workflows.

NVIDIA Nemotron 3 Ultra is an open frontier-reasoning and orchestration model featuring an extensive 1,000,000-token context length. Built on a hybrid Transformer-Mamba mixture-of-experts (MoE) architecture, it operates with 55B active parameters out of a total of 550B parameters. This advanced structure is designed to support sophisticated reasoning and complex task orchestration.

Developed by NVIDIA, Nemotron 3.5 Content Safety is a compact 4B-parameter multimodal guardrail model fine-tuned from Google Gemma-3-4B. Supporting an expansive context length of 128,000 tokens, this model moderates both user inputs and generated responses for LLMs and VLMs. It provides robust safety filtering to ensure aligned and secure interactions across multiple modalities.