NVIDIA squeezed a 74B-parameter open model into 24GB

NVIDIA released Nemotron 3.5 Bolt on Oct 15: an open-weights model with 74B parameters (9B active per token) that stores most of its weights at about 2 bits to fit in 24GB of memory, under a license that allows commercial use. NVIDIA's own llama.cpp recipes ask for a 32GB+ GPU such as the RTX 5090. In NVIDIA's own tests it beats Qwen3.6-35B-A3B on web research, but it trails rivals on factual recall.