Nvidia open-sources Nemotron 3.5 Bolt 74B-A9B agent model

Nvidia announced Nemotron 3.5 Bolt 74B-A9B on October 13, an open agent model with 74B total parameters and roughly 9B active per token, compressed to about 24GB via 2-bit quantization. Weights, training data and recipes are published; the release was scheduled for October 15, a day before RTX Spark laptops go on sale.

• 74B total parameters, about 9B active per token in a sparse MoE design
• 2-bit quantization-aware training brings weights to roughly 24GB
• Supported runtimes include llama.cpp, vLLM, Ollama, Unsloth and LM Studio
• No independent benchmarks yet; weights unconfirmed at publication time