What NVIDIA Nemotron 3.5 Super VL 120B-A12B is
NVIDIA Nemotron 3.5 Super VL 120B-A12B is an open-weight multimodal model from NVIDIA. It belongs to the Nemotron 3.5 Super family and is aimed at workloads that combine language understanding with visual analysis. The model can process text, images, sampled video frames, and document pages represented as images with accompanying text. Its output is text rather than generated images, video, audio, or speech.
The model has 120 billion total parameters, with approximately 12 billion parameters active for each token through a mixture-of-experts, or MoE, design. In practical terms, this allows the model to have a very large overall capacity without activating every parameter for every part of a request. It is still a substantial model to deploy, particularly in its BF16 form, so its architecture should not be interpreted as making it suitable for low-memory hardware.
NVIDIA released the model on October 15, 2026. The canonical model identifier is nvidia/nemotron-3.5-super-vl-120b-a12b. NVIDIA makes it available through Build.NVIDIA.com and NVIDIA NIM, while public BF16 and NVFP4 checkpoint variants are available through NVIDIA's Hugging Face organization.
Architecture and context window
Nemotron 3.5 Super VL combines interleaved Mamba-2 layers, Transformer attention, and mixture-of-experts layers. Mamba-2 is a sequence-modeling architecture intended to process long sequences efficiently, while Transformer attention helps the model relate information across a request. The model also uses NVIDIA's C-RADIO vision encoder to turn visual information into representations that can be processed alongside text.
A major specification is its context length of up to 1,048,576 tokens. Context is the working space available for the input and conversation, so this limit can accommodate very large collections of text, long documents, or extensive multimodal evidence when the deployment configuration supports it. A one-million-token limit does not mean every request will be equally fast or inexpensive: the required hardware, visual preprocessing, reasoning setting, and generated output all affect operational cost and latency.
The model supports native multi-token prediction. This technique can help speculative decoding systems predict multiple tokens during an inference step, potentially improving generation efficiency when the serving stack is configured to use it. NVIDIA also provides model-specific settings for video decoding, multimodal processing, reasoning, tool calls, and speculative decoding through NIM.
Inputs, outputs, and reasoning behavior
The verified input modalities are text, images, and video. Video requests can be handled through sampled frames, allowing the model to analyze visual content over time rather than treating a request as a single still image. Document analysis can combine page images with text, which is useful for layouts, charts, forms, and other information that plain text extraction may lose.
Audio input is not supported by the documented BF16 and NVFP4 checkpoints. The model also does not provide native image, video, audio, music, or speech output. Its response is text, including explanations, extracted information, code, classifications, and tool-call arguments where the serving interface supports them.
Reasoning is configurable. Users can enable or disable thinking and select different reasoning-effort levels. This creates a practical trade-off: more reasoning may help with difficult visual questions, multi-step analysis, coding, or tool-assisted tasks, but it can increase latency and token usage. Video requests may require higher output-token limits when reasoning is enabled, although NVIDIA has not published a verified maximum output-token limit for this exact model in the supplied documentation.
NVIDIA NIM supports structured machine-readable responses. JSON-object output is intended to produce valid JSON, while JSON Schema structured output is the more appropriate option when a response must contain specific required fields. These guarantees depend on the NIM interface and configuration; they should not automatically be assumed for every possible local checkpoint integration.
Coding, tools, and agentic workflows
Nemotron 3.5 Super VL is positioned for coding as well as multimodal reasoning. It can be useful when code generation or debugging depends on visual evidence, such as a screenshot of an interface, a diagram, a technical document, or a recorded workflow. It can also analyze documentation and code-related material within a long context window.
The model supports tool calling through NVIDIA's NIM interface. This allows an application to provide functions or external operations that the model can request, such as searching a data source, retrieving a document, running a business action, or passing information to another service. The model does not independently perform those operations: the surrounding application must define the tools, execute approved calls, and return their results.
These features make the model suitable for agentic systems that repeatedly inspect information, reason about the next step, call a tool, and produce a final text response. The long context can help such systems retain documents, prior observations, and intermediate results, but developers still need controls for tool permissions, validation, timeouts, and potentially sensitive visual data.
Deployment options and availability
NVIDIA provides hosted access through Build.NVIDIA.com and deployment support through NVIDIA NIM. NIM exposes OpenAI-compatible and Anthropic-compatible interfaces, along with streaming and configuration options specific to multimodal processing, video decoding, reasoning, tool calls, and speculative decoding. This can reduce integration work for applications already designed around compatible APIs, but the exact behavior remains dependent on the selected NIM deployment and configuration.
For self-managed deployments, NVIDIA publishes BF16 and NVFP4 checkpoint variants. BF16 is a higher-precision format commonly used when preserving model quality is the priority, while NVFP4 is intended to reduce memory requirements for optimized NVIDIA inference. Neither option makes the model a lightweight deployment: the total model size, visual encoder, runtime overhead, context length, and concurrency requirements all matter when planning hardware.
Fine-tuning is supported for the BF16 checkpoint, including documented LoRA workflows. LoRA, or Low-Rank Adaptation, can adapt selected model behavior without updating every model parameter. This may be useful for organizations with specialized document formats, visual procedures, or domain-specific response requirements, but it still requires suitable training and inference infrastructure.
Strengths and limitations
- Long context: The maximum documented context length is 1,048,576 tokens, making the model suitable for large document collections and extended multimodal tasks.
- Multimodal understanding: Text, images, and video are handled by one model, supporting document, visual, and temporal analysis workflows.
- Agent support: Tool calling, streaming, configurable reasoning, and structured response options are available through NIM.
- Deployment flexibility: NVIDIA offers hosted and NIM-based access, alongside BF16 and NVFP4 open checkpoint variants.
- Large infrastructure requirement: A 120B-parameter MoE model remains demanding to run, especially for BF16 inference, long contexts, or multiple simultaneous users.
- Text-only output: The model interprets visual inputs but does not generate native images, video, audio, or speech.
- No audio input: Applications centered on spoken conversations or audio analysis need a different model or an additional audio-processing system.
- Unverified serving limits: No official public per-token hosted price, maximum output-token limit, prompt-caching specification, or batch API specification was verified for this exact model.
The editorial ratings supplied for reasoning, coding, speed, and cost are comparative estimates rather than NVIDIA-published scores. They suggest a model optimized more for capability and complex multimodal work than for minimal latency or minimal infrastructure expense. Actual speed and cost will vary substantially with hardware, quantization, context size, video processing, concurrency, and reasoning settings.
Pricing and operational cost
No verified public per-token price was supplied for Nemotron 3.5 Super VL 120B-A12B, so there is no reliable price amount or billing period to report. Open-weight availability does not mean that operation is free: self-hosting involves GPU or accelerator capacity, storage, networking, engineering, and maintenance costs, while NIM or hosted access may have provider-specific infrastructure or usage charges.
For budgeting, the most important variables are likely to be the selected checkpoint, hardware format, number of concurrent requests, input context size, number of video frames, generated reasoning tokens, and whether speculative decoding is enabled. These factors make a simple comparison based only on parameter count or a nominal model price incomplete.
Best use cases
This model is a strong fit when a system must reason over large amounts of visual and textual evidence and then return a text answer or take a tool-assisted action. Suitable examples include:
- Question answering over long technical documents, scanned pages, charts, and diagrams.
- Video understanding based on sampled frames, including event review and visual investigation.
- Multimodal coding assistance involving screenshots, documentation, interface states, or technical diagrams.
- Long-context research and analysis where evidence must remain available across a large request.
- Agents that inspect documents or images, call approved tools, and return structured results.
- Specialized applications adapted with LoRA using the BF16 checkpoint.
When to choose this model
Choose Nemotron 3.5 Super VL when multimodal understanding, long context, configurable reasoning, and open-weight deployment matter more than the smallest possible serving footprint. It is particularly compelling for teams that want one model to combine document, image, video, coding, and tool-use tasks, or that need to deploy through NVIDIA's ecosystem and tune a BF16 checkpoint.
A smaller or text-only model may be more appropriate for short conversational requests, high-volume low-latency workloads, or deployments where GPU memory and operating cost are tightly constrained. A dedicated audio-capable model is preferable when speech or sound is central to the application. A model with native media generation should be selected when the required output is an image, video, audio clip, or speech rather than text.
Nemotron 3.5 Super VL is therefore best understood as a high-capability multimodal reasoning component, not a universal media model and not a lightweight chatbot. Its value comes from combining long-context visual understanding with coding and agent tools; its main trade-offs are infrastructure demands, uncertain hosted pricing, text-only output, and the absence of audio input.
Availability and model identifiers
The canonical identifier is nvidia/nemotron-3.5-super-vl-120b-a12b. Public checkpoint variants include NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16 and NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4. NVIDIA's documented release channels include Build.NVIDIA.com, NVIDIA NIM, and the NVIDIA Hugging Face organization. These identifiers should be checked carefully when configuring a serving system because the checkpoint format and deployment path can affect supported features and resource requirements.