← All models

Nemotron 3.5 Bolt 74B A9B BF16

NVIDIA · 2026-10-15 · 74.9B parameters

Nemotron 3.5 Bolt 74B A9B BF16 is the full-precision release of NVIDIA's Nemotron 3.5 Bolt, a hybrid Mamba-2, attention and LatentMoE vision-language model compressed 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 tokens of context and optional thinking. NVIDIA positions this BF16 checkpoint (74,885,081,096 parameters) for customization and post-training, and recommends the separately released 2-bit 24GB build for inference; the 24GB build and its GGUF are distinct artifacts with their own scores.

Benchmark scores