NVIDIA's Nemotron-3.5 Bolt targets 24GB GPUs

NVIDIA has published Nemotron-3.5 Bolt, a multimodal mixture-of-experts model that pairs a large 74-billion-parameter capacity with a lean nine-billion active footprint. The release is available now on Hugging Face.

The naming is unusually explicit about its intended home: the "24GB" suffix signals that NVIDIA has sized this model to run on a single 24GB GPU, the class of memory found in cards like the RTX 4090 and professional workstation parts. That framing matters because large vision-language models have typically demanded multi-GPU setups or datacenter accelerators.

Why the MoE math matters

Mixture-of-experts designs route each token through only a fraction of the network, so the compute and memory cost at inference tracks the active parameters rather than the full total. Here that means:

• 74B total parameters for overall capacity
• 9B active parameters per forward pass
• A target of 24GB of GPU memory

That combination is the central pitch — expert-level breadth without the full hardware bill. The model is listed as a vision-language model, extending Nemotron beyond text into multimodal understanding.

NVIDIA has not published a context length in this record, and the model carries a custom "other" license, so teams will want to review the terms before building on it. Still, a multimodal MoE engineered for consumer-class memory lowers the barrier for developers who want capable vision-language inference on hardware they already own.