Segmenting models by deployment location
NVIDIA released Nemotron 3.5 Super VL and Bolt 74B-A9B 24GB on October 15, 2026. Super VL targets visual documents and multistep agent work; Bolt targets local workloads on RTX and DGX Spark systems. The smaller Bolt E4B 4GB was scheduled for October 23 in the announcement and should not be presented as already downloadable.
We verified Bolt GGUF weights in NVIDIA's official Hugging Face repository. NVIDIA describes native 2-bit quantization designed to fit 24GB, but that label is not a guarantee that total runtime memory stays below 24GB for every context, batch, or operating system. KV cache and runtime overhead still matter.
Open weights do not eliminate operating costs
The release expands options for local-agent users and organizations with data-residency requirements. Vendor speed and accuracy comparisons remain specific to their test conditions. Downloadable weights do not remove GPU, power, or support costs, and each checkpoint's license still needs review. A small pilot should check runtime compatibility, long-input memory, and successful tool use before a production commitment.