Nemotron 3.5 Bolt

NVIDIA Nemotron 3.5 Bolt 24GB is designed for executing agentic tasks directly on NVIDIA RTX and DGX Spark hardware. This model is a compact version of Nemotron 3.5 Super VL, compressed from 120B total/12B active parameters to 74B total/9B active, specifically engineered to fit within a 24GB memory budget. It features both text and vision capabilities, making it versatile for various applications requiring multimodal processing. NVIDIA reports an impressive decoding speed of 84 tokens per second on RTX Spark, ensuring rapid task execution. The model utilizes advanced compression techniques, including Star Elastic, to maintain high performance despite its reduced size. Post-training and quantization-aware distillation further optimize its capabilities. Checkpoints, GGUF files, and a speculative-decoding drafter are available on Hugging Face, providing developers with the necessary resources for deployment and integration. This robust solution is ideal for developers and engineering teams requiring powerful, efficient processing for local agentic workflows on NVIDIA hardware, offering a strong balance of processing speed and functional accuracy for memory-constrained environments.