NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB

Overview

NVIDIA's 74B (9B active) hybrid Mamba-2 MoE vision-language model, quantization-aware-trained to 2-bit so the whole model packs into ~24 GB - Terminal-Bench 2.1 at 51.0, ahead of everything in its weight class. Handles text, images and video with a 256K context window.

Good for

• agentic
• chat
• coding

Recommended hardware

≈ 32 GB VRAM to run this 74B model at the IQ2 quantization.

Needs a 32 GB+ GPU (e.g. RTX 5090, or multi-GPU / 2x24 GB)

Links

• View on HuggingFace
• Run on Infersec