NVIDIA released Nemotron 3.5 Bolt 74B-A9B, a 2-bit vision-language model that fits in 24 GB

Archive entry: this covers an event from October 15, 2026, and we published it on November 8, 2026. The details come from sources dated at the time and may have changed since.

On October 15, 2026, NVIDIA released NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB on Hugging Face, a vision-language model for local everyday agentic tasks. According to the model card on Hugging Face, native 2-bit weights from quantization-aware training let the 74B-parameter model fit in 24 GB.

What changed

• Native 2-bit IQ2 weights from quantization-aware training let the model fit in 24 GB.
• It had 74B total parameters with 9B active, in a Mamba-2-Transformer hybrid LatentMoE with a C-RADIO v4-H vision encoder.
• It had a 256K context and accepted text, image and video input.
• It shipped with a LiLiCorr DFlash drafter for 7 speculative tokens.

What it meant for people building with AI

Developers with a single 24 GB GPU could run a 74B-parameter multimodal agent model locally under a commercial-use license. The memory fit came from the native 2-bit IQ2 weights, and the model took text, image and video input with a 256K context, so one local model covered several kinds of input.

Runtime, license and release date

The model card said the model ran with vLLM on A100, H100-80GB and Blackwell GPUs. The license was OpenMDW-1.1, which allows commercial use, and a GGUF version existed. The card gave the Hugging Face release date as October 15, 2026.

Sources

• nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB (NVIDIA, October 15, 2026, official)

Written with AI assistance, strictly from the sources listed above. Before it went live, automated checks verified its dates, sources and every language edition.