Description
NVIDIA Nemotron 3.5 Super VL is a sophisticated large language model hosted on Hugging Face. It is designed to handle a variety of inputs, including text, images, and video, and outputs primarily in text format. The model is built on a Mamba-2–Transformer hybrid architecture, incorporating a C-RADIO vision encoder, which allows it to process multimodal data efficiently.
Read More
With a total of 120 billion parameters, of which 12 billion are active, Nemotron 3.5 Super VL is engineered for high performance in tasks requiring customization and domain adaptation. It supports a context length of up to 1 million tokens, making it suitable for extensive data processing tasks. The model operates with BF16 precision, ensuring high accuracy in its outputs.
Deployment of Nemotron 3.5 Super VL can be achieved using either 4× H100 or 2× GB200 hardware configurations, allowing flexibility based on available resources. The model supports speculative decoding with native multi-token prediction, enhancing its efficiency in generating outputs.
Nemotron 3.5 Super VL is particularly recommended for tasks that require full-precision evaluation. It offers configurable reasoning modes through chat templates, allowing users to enable or disable thinking modes as needed. The recommended sampling settings for text include a temperature of 1.0 and top_p of 0.95, while for vision tasks, a temperature of 0.6 and top_p of 0.95 are suggested.
The model is licensed under OpenMDW-1.1, promoting open-source use and collaboration. It has been benchmarked against several other models, demonstrating competitive performance across various tasks, including GPQA-Diamond, HLE, and MMLU-ProX. These benchmarks highlight its capability in handling complex reasoning and data processing tasks.
Overall, NVIDIA Nemotron 3.5 Super VL is a versatile tool for developers and researchers looking to leverage advanced AI capabilities for multimodal data processing and analysis.
- CategoryAI Models & LLMs
- Deployment4× H100 or 2× GB200
- PlatformsWeb
NVIDIA Nemotron 3.5 Super VL Highlights
Mamba-2–Transformer hybrid architecture
C-RADIO vision encoder
120 billion parameters
BF16 precision
Context length up to 1 million tokens
Speculative decoding with multi-token prediction
Configurable reasoning modes
OpenMDW-1.1 license
Getting Started with NVIDIA Nemotron 3.5 Super VL
Access model page: Visit Hugging Face
Authenticate: Log in to your account
Configure: Set up hardware and parameters
Send prompts: Input text, image, or video data
Fine-tune: Adjust settings for specific tasks
NVIDIA Nemotron 3.5 Super VL's Use Cases
- Text Processing
- Image Analysis
- Video Processing
- Domain Adaptation
- Full-Precision Evaluation
