Muse Spark Open
Meta’s open-weight multimodal model supports coding, tool workflows and long-context applications through self-hosting or hosted inference.
NVIDIA Nemotron 3.5 Super VL
research.nvidia.com
NVIDIA’s open-weight vision-language model supports coding, visual reasoning, document understanding and agents that use tools.
Visit website
Pricing
Free
Available on
Linux, NVIDIA GPU, Self-hosted, Hosted API
Category
Llm Models
What is NVIDIA Nemotron 3.5 Super VL?
NVIDIA Nemotron 3.5 Super VL is an open-weight multimodal model with 120 billion total parameters and 12 billion active parameters. It supports reasoning, coding and interactive tool workflows across text, images and video. NVIDIA releases BF16 and NVFP4 checkpoints with inference recipes.
The model serves developers building document-understanding systems, visual reasoning applications and multimodal agents. NVIDIA’s usage cookbook provides deployment, fine-tuning and agent examples. The NVFP4 repository supplies a quantized checkpoint for deployment.
Key features
Visual and video understanding
Interpret images, document-page images and video; available input types depend on the deployment recipe.
Long-context support
The released model supports context lengths up to one million tokens for long-context understanding and agent workloads.
Configurable reasoning
Choose reasoning-off, regular or medium-effort operation, with inference-time controls for the reasoning budget.
Inference optimization
BF16 and NVFP4 releases have optimized serving recipes. Native Multi-Token Prediction supports speculative decoding, with configurations tuned to hardware and workload.
Adaptation and agent recipes
Developer cookbooks cover document-image LoRA fine-tuning, multimodal reinforcement learning and a visual investigation agent.
Who is it for?
Developers building document question answering
Create a service that answers questions about scanned pages, charts and accompanying text.
Software teams developing coding agents
Connect repository and execution tools to support code investigation and proposed changes.
Teams inspecting visual interfaces
Build an agent that interprets dashboard screenshots and selects relevant application tools.
Researchers adapting multimodal models
Use the supplied fine-tuning recipes to explore a specialized document-image task.
How to use NVIDIA Nemotron 3.5 Super VL
- Review the model license and choose a checkpoint and deployment recipe for your application.
- Match hardware, runtime and input types. For NVFP4 video, consult the checkpoint deployment table; vLLM on one H200 is a documented option.
- Start the server using the selected recipe and wait for model loading and kernel compilation to finish.
- Check endpoint readiness, then send a text or supported visual request through a compatible API client.
- Use the cookbook’s agent or fine-tuning examples when the application needs connected tools or task adaptation.
NVIDIA Nemotron 3.5 Super VL pricing
Free
Downloadable model materials are licensed without charge under OpenMDW-1.1, subject to its terms. Infrastructure and hosted inference cost separately. NIM offers free development access, while production NIM licensing is based on GPU count. Together AI also lists paid token-based inference.
Check current pricing ↗What to know before you use it
- The documented stack uses Linux and NVIDIA GPUs. For NVFP4, listed SGLang and DGX Spark vLLM recipes support text and images; vLLM on two H100s, one H200 or one GB200 also supports video. Consult the checkpoint deployment table.
- Documented visual formats include JPEG, PNG and MP4 video up to two minutes. Documents use page images with accompanying text.
- Inputs and generated output share the configured context window; deployments may configure a smaller window.
- Free NIM Developer Program access covers prototyping, research, development and testing. Production use of the NIM runtime requires NVIDIA AI Enterprise licensing, separate from the model license.
Frequently asked questions
Does the model generate images or video?
Its output is text; images and video are inputs.
What does the NVFP4 checkpoint provide?
It provides quantized weights with lower memory requirements than BF16.
Is NVIDIA AI Enterprise required for every deployment?
Production use of NIM requires NVIDIA AI Enterprise. The underlying model has a separate license, and NVIDIA documents other inference runtimes.