Choose theme
AI Resources
NVIDIA Nemotron 3.5 Super VL Base
NVIDIA Nemotron 3.5 Super VL Base is a downloadable vision-language checkpoint for building custom models through further training.
This page covers the 120B-A12B-Base-BF16 release. It takes text, images, and sampled video frames and produces text. NVIDIA publishes separate post-trained checkpoints for readers starting from a model already adapted to assistant tasks. Use this as a first read, not a recommendation. Open the original project before trusting details like terms, limits, privacy, cost, setup, or safety.
What it is
A starting point for further training
A base model for developers and researchers who want to continue training a model that works with language and visual inputs.
Why it stands out
Language and vision in one context
Its hybrid language backbone connects to a vision encoder, with a documented context ceiling of one million tokens shared by text and visual inputs.
Availability
Downloadable BF16 checkpoint
The BF16 weights, configuration, and model code are on Hugging Face. The model card names OpenMDW License Agreement, version 1.1; use that main page for current terms.
Why it matters
What makes it useful
A team adapting AI to its own documents can start with a checkpoint that already connects page images and language. The model card accepts documents as page images with accompanying text, so further training can include the visual page rather than relying only on extracted words.
What to know
Where it fits
This sits at the model-training stage: NVIDIA lists continued pre-training, supervised fine-tuning, reinforcement learning, and distillation as intended uses. The Base checkpoint leaves that adaptation work with the builder. NVIDIA links a post-trained BF16 sibling for customization and an NVFP4 sibling for deployment from the same model card.
Notable points
What stands out
The one-million-token ceiling is a shared budget: the prompt, images or video frames, conversation history, tool results, and generated answer all use it. A serving recipe may choose a smaller window to fit GPU memory and concurrent requests, so the headline length does not tell you how many document pages a deployment will handle.
Before using
What to review
Match the exact Base BF16 checkpoint to your training framework and GPU-memory plan. The 12B active figure describes the sparse computation, while the model card lists 120B parameters in total.
The benchmark table on the Base model card evaluates the text-only backbone before vision-language training. Those numbers do not establish this checkpoint's document or video performance.
NVIDIA specifies RGB JPEG/PNG images and sampled MP4 video frames, with video input up to two minutes. Check the current card and your chosen data-processing path for those limits.
Use task-specific data to evaluate the adapted model. NVIDIA calls for additional testing when integrating foundation and fine-tuned models into a system.
Reader fit
Who may find it relevant
Most relevant to model-training teams and researchers choosing a starting checkpoint for domain adaptation. A reader wanting an assistant to use immediately can follow the post-trained model link below; the Base release is a training artifact rather than a finished chat service.
Editorial note
Why LifeHubber lists it
We list this because NVIDIA exposes the vision-language foundation as its own checkpoint alongside the post-trained releases. For a team investigating how to specialize document or video understanding, that makes the starting stage explicit: it can consider further training from Base or customization from an assistant-trained sibling. Keeping those paths visible helps readers choose what they want to adapt before choosing how to serve it.
Source links
Source materials
Reader note
Before relying on this entry
LifeHubber lists entries to help readers inspect AI projects, not to endorse them or prove they are safe, suitable, accurate, maintained, or right for a specific use. We do not verify every entry in depth. Before relying on anything listed, review the original materials, terms, privacy practices, limits, and risks that matter for your situation.
More in AI Models
Keep browsing this category
Explore more AI model resources.
Gemma 4
google/gemma-4
A Google DeepMind Gemma 4 model family collection with public checkpoints including Gemma 4 12B, a dense multimodal model Google describes around local agentic workflows, native audio input, and encoder-free vision/audio handling.
MiniMax H3 Integrations
MiniMax-AI/awesome-minimax-h3-integration
A community-maintained MiniMax H3 integration index that maps checkpoints, hardware and VRAM starting points, runtimes, ComfyUI nodes, prompting tools, acceleration routes, and deployment options.
North Micro Vision Instruct
CohereLabs/North-Micro-Vision-Instruct
Cohere Labs' downloadable 2.4B vision-language model for native-resolution documents, charts, OCR, visual grounding, image questions, and task-specific fine-tuning, with multilingual and multi-image input plus Transformers, NVIDIA AutoModel, Axolotl, and community MLX paths.
For project maintainers
Listed here? You can use the badge.
If you maintain a project with a current LifeHubber listing, you may add the optional “Listed on LifeHubber AI Resources” badge to its README, docs, or website. No introduction or permission request is needed.
