Nemotron 3.5 Bolt 74B-A9B

NVIDIA Nemotron 3.5 Bolt 74B-A9B 24GB is a hybrid vision-language model designed for reasoning, multimodal understanding, and local agentic workloads. Derived from the larger Nemotron 3.5 Super VL 120B-A12B model family, it combines Star Elastic compression and Quantization-Aware Distillation to optimize intelligence and memory footprint. The model features 74 billion total parameters with 9 billion active per token, allowing it to fit into 24GB of memory for localized execution.

Architecture and Capabilities

The model utilizes a hybrid Mamba-2 and Transformer mixture-of-experts (MoE) architecture integrated with a vision encoder. This design enables high inference throughput and efficient handling of multimodal data, including images, screenshots, charts, documents, and video inputs. It is optimized for complex edge agent applications such as browser interaction, terminal operations, and structured tool calling.

Optimized for edge deployment on hardware like NVIDIA RTX Spark and DGX Spark, the model leverages native low-bit quantization (such as IQ2_XXS expert weights) alongside auxiliary layers to preserve strong multi-turn reasoning and coding performance.