Nvidia launches nemotron 3.5 super vl, a 120 billion parameter multimodal model with enhanced throughput and efficiency

NVIDIA has unveiled Nemotron 3.5 Super VL, a large multimodal model with 120 billion parameters, designed for continuous operation and multi-step tasks. The model, built on a mixed expert (MoE) architecture, achieves up to five times higher throughput than other open models on a single GPU. It integrates reasoning, instruction following, and tool calling with text, image, document, and video understanding, enabling applications like code testing, cross-site search, and cybersecurity threat detection. Developers can fine-tune the model using NVIDIA’s NeMo toolkit. NVIDIA highlighted several enterprise use cases, including a 97.5% reduction in inference costs for Scale AI’s credit processing agent and a 90% drop in latency for Ramp’s receipt processing. The model also outperformed previous versions in benchmark tests, achieving 80.8% on CharXiv and 84% with Python code interpretation. It supports speculative decoding, mixed precision, and NVFP4 quantization, with 4-bit weights and activations on Blackwell architecture. NVIDIA also introduced two Bolt models, with the 74B-A9B 24GB version now available and the E4B 4GB model set for release on October 23. Both are accessible via Hugging Face and other platforms, with deployment options ranging from DGX systems to cloud environments.