NVIDIA expands Nemotron 3.5: 120B Super VL and a 2-bit Bolt 74B-A9B that fits in 24 GB
In one sentence On October 15 NVIDIA released Nemotron 3.5 Super VL, a 120-billion-parameter open vision-language model, and Nemotron 3.5 Bolt 74B-A9B, natively quantized to 2 bits to fit in 24 GB of memory on RTX PCs and DGX Spark. A 4 GB Bolt E4B variant is announced for October 23.
NVIDIA keeps widening its Nemotron family of open models. The first newcomer, Super VL, understands text and images with a clear strength: reading documents, tables, charts, web pages and PDFs. NVIDIA publishes not only the weights but also datasets and training recipes.
The second, Bolt 74B-A9B, is built to run locally: it has 74 billion parameters but uses 9 at a time and is compressed to 2 bits from the start, so it fits in 24 GB of memory. That means it can run on a high-end GeForce RTX card or a DGX Spark while leaving room for other applications.
Both are already on Hugging Face, and Bolt works with the most common local AI tools such as llama.cpp, Ollama and LM Studio. A smaller 4 GB sibling for voice assistants and interactive characters is expected on October 23.
Companies
NVIDIA
Tools
Nemotron 3.5 Super VL, Nemotron 3.5 Bolt 74B-A9B, Nemotron 3.5 Bolt E4B
Tags
Sources