Nemotron 3.5 Bolt 74B-A9B (precision unspecified)
NVIDIA · 2026-10-15
NVIDIA Nemotron 3.5 Bolt 74B-A9B labels whose evaluators do not identify the served BF16 or 2-bit 24GB artifact. NVIDIA released three builds on Hugging Face on October 15, 2026: the BF16 checkpoint (nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-BF16, 74,885,081,096 parameters), which it positions for customization and post-training; a 24GB build whose experts are quantized to IQ2_XXS (about 2.06 bits) with quantization-aware distillation, which it recommends for inference; and a GGUF of that 24GB build. NVIDIA's own model cards report different scores for the BF16 and 24GB builds. Bolt is a hybrid Mamba-2, attention and LatentMoE vision-language model compressed with Star Elastic from Nemotron 3.5 Super VL to about 74B total and 9B active parameters, with a C-RADIO v4-H vision encoder, text, image and video input, up to 256K-token context and optional thinking under OpenMDW-1.1. This benchmark entry preserves the source observations without assigning them to a checkpoint or including the unresolved identity in the capability fit.