NVIDIA's Nemotron 3.5 Bolt 74B-A9B 24GB is available now for local agents
On Oct 15, alongside Nemotron 3.5 Super VL, NVIDIA released Nemotron 3.5 Bolt 74B-A9B 24GB: an open vision-language model with 74B total parameters and 9B active, trained with native 2-bit quantization to fit in 24 GB. It takes text, images and video, writes text, supports up to 256K context, and is aimed at everyday local agent work on RTX/DGX Spark and other NVIDIA systems. Weights are on Hugging Face under OpenMDW-1.1; NVIDIA lists support for llama.cpp, Ollama, LM Studio, Unsloth and vLLM. Bolt E4B (4GB) is due Oct 23. Benchmark rankings are NVIDIA's own claims. (NVIDIA)
How you’d use it: If you want an open local agent model that fits in 24 GB of GPU memory, download Bolt from Hugging Face and try it with llama.cpp, Ollama, LM Studio or vLLM. Treat the leaderboard numbers as NVIDIA's own.