NVIDIA released Nemotron 3.5 Super VL and Nemotron 3.5 Bolt 74B-A9B on October 15, expanding its downloadable model range for teams building agents that read documents, call tools and complete work across multiple steps. A third model, Bolt E4B 4GB, is scheduled for October 23 and is not part of the immediately available pair.
Super VL has 120 billion total parameters and activates 12 billion per token. It combines text and visual understanding, so an agent can use information from images, charts, document pages and video while handling a task. NVIDIA positions it for workflows such as document processing and enterprise research, with post-training tools for adapting it to an organization's own data and procedures.
Two different memory budgets
The Super VL model card lists a context window of up to one million tokens. Prompts, visual inputs, tool results, conversation history and the answer share that space. A deployment can use a smaller window to accommodate its available memory and concurrent requests.
NVIDIA supplies NVFP4 deployment recipes for hardware including one DGX Spark, one H200 or two H100 GPUs, with different supported input configurations. The weights are governed by the OpenMDW License Agreement 1.1. Downloadable weights do not remove hardware, hosting or integration costs.
Bolt takes a different route. NVIDIA compressed the larger Super VL model to 74 billion total and nine billion active parameters, then used aggressive quantization to target a 24GB footprint. Most expert weights use approximately two bits per element; other parts of the network retain different precisions. Describing the whole model as uniformly two-bit would miss that distinction.
The release includes a quantized checkpoint, an unquantized BF16 version for customization and a separate speculative-decoding model. The quantized version is optimized for RTX Spark with llama.cpp and DGX Spark with vLLM.
Speed claims need their test conditions
NVIDIA's technical material reports roughly 84 output tokens per second for Bolt on an RTX Spark Surface Laptop Ultra with 128GB of memory, using its specified decoding setup. In the same table, the tested Qwen3.6 and Gemma 4 configurations generate tokens faster and finish the replay benchmark sooner. NVIDIA's argument is a balance of capability and memory use, not a universal speed win over every smaller model.
For Super VL, the company separately advertises up to five times higher throughput than selected open models in its class. Its technical blog distinguishes that output-speed comparison from an agent task-completion test. Neither figure establishes that an arbitrary production workflow will finish five times faster. The model card explicitly says performance depends on hardware, request length and concurrency.