Single model for reading, seeing, and acting, NVIDIA introduces Nemotron 3.5 Super VL

NVIDIA unveils Nemotron 3.5 Super VL, an open 120-billion-parameter multimodal model designed for enterprise AI agents.

Reading a contract, interpreting a chart, examining a video sequence, and then using software to complete a task. These operations sometimes require several specialized models, with information exchanges between each step. NVIDIA Nemotron 3.5 Super VL seeks to simplify this setup by combining visual understanding, reasoning, and tool use into a single model. With 120 billion parameters, of which only 12 billion are active at each step, NVIDIA prioritizes an architecture suited for agents capable of working on extended professional tasks. A multimodal model to avoid switching between multiple AIs

Nemotron 3.5 Super VL supports text, images, documents, and videos. It can analyze a page containing tables, cross-reference multiple charts, or leverage visual information in broader research.

The value does not lie solely in recognizing displayed elements. The model can link these observations to instructions, request tools, and continue a multi-step task.

An agent could, for example, examine scanned invoices, extract amounts and dates, compare the information with accounting software, and then flag inconsistencies. The same architecture can be adapted to technical documentation, incident analysis, or information retrieval in complex files.

NVIDIA thus seeks to limit context transfers between a language model and a vision model. This integration simplifies certain processing pipelines without eliminating the use of external services when the task requires it. 120 billion parameters, with 12 billion active

Nemotron 3.5 Super VL relies on a hybrid architecture combining Mamba-2, Transformer, and Mixture-of-Experts (MoE).

Out of its 120 billion parameters, approximately 12 billion are mobilized for each processed token. This design provides significant capacity while limiting the operations required for each generation.

The model relies on a C-RADIO visual encoder, responsible for converting image and video information into representations usable by its language component.

Its context window can reach one million tokens, including instructions, documents, images, tool results, and interaction history.

The documentation notably provides support for MP4 videos up to two minutes long, based on frames extracted from the sequence. This involves video understanding, not generating new videos.

Actual context limits may be lower than the announced maximum depending on available memory, runtime settings, and the number of concurrent requests. Up to four times faster on 10,000 programming tasks

NVIDIA particularly highlights the model's speed in agentic workflows.

In an evaluation covering the execution of 10,000 programming tasks on NVIDIA B200 GPUs, Nemotron 3.5 Super VL reportedly completed all the work up to four times faster than certain comparable open models, while maintaining strong performance on SWE-bench Pro.

This metric reflects the time required to complete tasks under defined conditions, not a guaranteed speedup for every request.

On Artificial Analysis, NVIDIA also reports speeds exceeding 600 output tokens per second on a Baseten endpoint, in a configuration using 100,000 input tokens.

The laboratory further states that it achieves up to five times more throughput on a single GPU compared to certain open models in its class.

These results depend on the hardware, the inference provider, and the workload. They should not be interpreted as universal performance metrics. Adaptations for finance, cybersecurity, and documents

Nemotron 3.5 Super VL is designed to be customized using an organization's data and procedures.

NVIDIA showcases several enterprise experiments conducted with partners.

At Skydance Corporation, Scale AI adapted the model for the automated processing of accounts payable credit files. For this application, NVIDIA announces an expected 40-fold reduction in inference cost and a five-fold improvement in response time.

Financial platform Ramp is working on converting photographs of receipts into structured data, notably including merchants, dates, and amounts. The company anticipates an increase of up to 50% in throughput and a tenfold reduction in latency.

For its part, AI21 customized the model for Alice, a cybersecurity company. NVIDIA reports a 40% accuracy improvement on threat and vulnerability analysis tasks.

These examples demonstrate the value of training adapted to a specific domain. However, some gains are presented as goals or deployment-specific estimates, rather than generalizable results for all enterprises. Reading charts and leveraging complex documents

Document analysis is one of the areas NVIDIA has focused on heavily.

The model can process multi-page documents, cross-reference information across different tables, and interpret the relationships between values and their graphical presentation.

On CharXiv, an evaluation dedicated to chart understanding, NVIDIA claims a score of 80.8%, compared to 63% for Nemotron 3 Nano Omni. With an additional interpretation step using Python, the score reaches 84%.

This performance relies on specific evaluation setups, but illustrates the progress pursued in transforming visual information into data usable by agents.

The model is also intended for video analysis, particularly for clip search, content examination, or incident investigation. Optimizations to reduce execution costs

The design of Nemotron 3.5 Super VL also relies on several methods aimed at reducing the resources needed for its operation.

The model notably leverages multi-token prediction, which allows it to propose several tokens before verification, rather than systematically processing each generation in isolation.

NVIDIA also presents LiLiCorr, a speculative decoding method that, in certain DGX Spark setups, enables generation up to 2.3 times faster than without speculation.

The NVFP4 quantized version reduces memory requirements through a more compact numerical representation, accompanied by training designed to limit precision loss.

Finally, context reuse mechanisms avoid unnecessarily