NVIDIA says its free Nemotron 3.5 Bolt E4B model runs in about 4GB

NVIDIA released Nemotron 3.5 Bolt E4B, a free, open-weights AI model built for fast local voice assistants, AI characters and devices, which NVIDIA post-trained from Google's Gemma 4 E4B. It has 8 billion parameters, 4 billion of them effective, and NVIDIA's model card says it's ready for commercial or non-commercial use. NVIDIA says it can run in as little as about 4GB: in NVIDIA's own tests it used 3.1 to 4.3 GB of graphics memory on RTX GPUs and hit 133 tokens a second on an RTX 5070. The catch: 4GB is memory use, not the download (the model file is 6.18 GB and Ollama's version is a 7.2 to 7.9 GB download), and in NVIDIA's own benchmarks the rival Qwen3.5 4B still wins on long-document tests.

ollama run nemotron-3.5-bolt:e4b