NVIDIA has released Nemotron 3.5 Super VL and Nemotron 3.5 Bolt 74B-A9B 24GB. Its launch announcement positions the models for different deployment settings, with a further Bolt E4B 4GB model scheduled for October 23.
Match the model to the deployment
Super VL is a 120-billion-parameter model for multimodal agent workflows. NVIDIA describes tasks involving documents, charts, tables, webpages and images, alongside instruction following and tool use.
Bolt 74B-A9B uses native 2-bit quantization to fit in 24GB of memory. The publisher targets local agent work on supported NVIDIA systems, including desktop research and coding workflows. The smaller E4B model is aimed at physical and conversational agents.
These are separate models with different resource goals. A memory claim about Bolt should not be applied to the larger Super VL model.
Evaluate the performance claims on your task
NVIDIA reports throughput advantages and specialized task results for Super VL. Its claim of up to five times higher throughput compares models within the class and evaluation described in the announcement.
The company also lists organizations customizing the model for particular workflows. Those examples illustrate its post-training pitch, but do not establish that a downloaded checkpoint will reproduce a partner's adapted result on unrelated tasks.
A useful evaluation should hold the task, serving conditions and quality criteria steady before comparing speed.
Check the available artifacts and serving options
The release includes open weights, datasets and training recipes. NVIDIA describes customization through NeMo and Python-based pipelines, with model access through Hugging Face, OpenRouter, ModelScope and NVIDIA NIM.
Those routes offer different operational choices. Inspect the model-specific requirements, applicable terms and supported inference configuration before selecting a deployment.
The immediate practical distinction is availability: Super VL and Bolt 74B-A9B are available now, while the announcement dates Bolt E4B for October 23. Treat the latter as a scheduled release until its artifacts and documentation are available.