NVIDIA has added three new options to the Nemotron open model family. The company’s October 15 announcement includes the Nemotron 3.5 Super VL for enterprise duties and two Bolt models for local use. with SuperVL Bolt 74B-A9B 24GB While it becomes available today, the smaller Bolt E4B 4GB model will arrive on October 23.
Super VL works as a 120 billion parameter model that understands images and documents along with text. NVIDIA is positioning the model for constantly running enterprise AI agents. In addition, companies can perform additional training with their own data and business rules and use the model in their own infrastructure.
The new family doesn’t stick to a single computer class. Super VL works in data center, cloud and appropriate local systems. Bolt 74B, on the other hand, focuses on running daily agent work on the computer. On the other hand, the Bolt E4B carries a smaller memory target for fast voice assistants, interactive characters and embedded interfaces.
NVIDIA emphasizes Super VL’s ability to handle tables, graphics, and text together within the document. The model can relate a chart on one page to a table in another section. Thus, developers aim to reduce the need to transfer context between separate text and image models in such tasks.
Bolt fits the model with 74 billion parameters into 24 GB memory
Bolt’s technical research page separates the total number of parameters and the active parameters at each step. The name 74B-A9B refers to approximately 74 billion total parameters and 9 billion active parameters. NVIDIA derives this model from the larger Super VL with compression and additional training steps.
Additionally, the team uses approximately 2-bit quantization for most of the model’s weights. So the model can be used to work with other applications. 24GB memory meets its target. However, this number does not guarantee the total memory required for all tasks of the computer or the same speed under all conditions.
Researchers performed speed measurements with llama.cpp on a Surface Laptop Ultra with 128 GB memory. Bolt produced approximately 84 tokens per second in the AIPerf SPEED-Bench-Coding test. On the other hand, in the same table, Qwen3.6 35B-A3B and Gemma 4 26B-A4B reached higher token rates. NVIDIA describes the Bolt not as the fastest model in all metrics, but as a balance between memory, accuracy and speed.
Super VL’s developer document describes the expert mix approach that enables approximately 12 billion of the total 120 billion parameters in each token. Additionally, document understanding training covers multi-page documents, optical character recognition, and video. With NeMo tools, developers can incorporate their own domain data and company policies into additional training.
In the technical document, Ramp’s work on converting receipt images into regular data containing vendor, date and payment information is among the examples. Ramp expects higher throughput and low latency in this business. However, this expectation of the company does not mean that every user who downloads the model will get the same result. While Bolt’s quantized weights and accelerator mockup are available for download today, NVIDIA is also announcing support for different native runtime tools.