NVIDIA has expanded its open-weight model ecosystem with the introduction of the Nemotron 3.5 Super VL, a 120-billion-parameter model designed for complex, always-on enterprise agents. The model utilizes a mixture-of-experts architecture that activates 12 billion parameters per token, allowing it to balance high-capacity reasoning with inference efficiency. By integrating text, image, document, and video understanding into a single model, it eliminates the need for organizations to deploy separate vision-language systems or manage complex task handoffs between different models.
The Super VL model is optimized for specialized workflows such as cybersecurity threat triage, automated financial processing, and web research. According to NVIDIA, the model delivers up to five times higher throughput for agentic tasks compared to other open models in its size category. Organizations including Scale AI, Ramp, and AI21 have already utilized the model for post-training, reporting significant gains in operational metrics, such as 40 times lower inference costs and 40 percent higher accuracy in specific production environments.
Alongside the flagship model, NVIDIA introduced two edge-focused variants under the Nemotron 3.5 Bolt umbrella. The Nemotron 3.5 Bolt 74B-A9B is a 24GB model that uses 2-bit quantization to fit into local memory, enabling multistep reasoning and tool calling on edge hardware without cloud dependency. The smaller Nemotron 3.5 Bolt E4B is a 4GB model optimized for fast conversational agents and embedded interfaces, designed to run on hardware including GeForce RTX and Jetson platforms.
These models incorporate architectural features from the previously released Nemotron 3.5 Lightning, including a hybrid mamba transformer, multi-token prediction, and speculative decoding. Developers can customize these models using NVIDIA NeMo for supervised fine-tuning and low-rank adaptation. NVIDIA has also released open datasets on Hugging Face to support training for agentic tool-use, conversational data, and financial reasoning, with the models available for deployment via NVIDIA NIM microservices.