NVIDIA Releases Nemotron 3.5 Bolt 74B: A Local-Agent Model Packed Into 24GB

What happened

On October 13, 2026, NVIDIA announced Nemotron 3.5 Bolt 74B-A9B 24GB on its official blog, with release set for Thursday, October 15. It is a new model built for local agents on RTX Spark laptops.

The model uses a mixture-of-experts (MoE) architecture: 74 billion total parameters with only 9 billion active per token, packed into 24GB via 2-bit quantization with quantization-aware training. NVIDIA says it balances speed and intelligence — faster than most popular dense models, smarter than most popular MoE models with a similar memory footprint.

At launch it runs through llama.cpp, vLLM, Ollama, Unsloth and LM Studio, and plugs directly into agents such as Hermes and OpenClaw. Like other Nemotron models, Bolt 74B is fully open source — training data, weights and recipes are all public, so developers can fine-tune and customize their own agents.

Also on the day: NVIDIA PAIR, a personal AI router, moved out of beta and went open source with new llama.cpp support; RTX Spark laptops go on sale October 16; the smaller Bolt E4B 4GB model (for ACE game characters) ships October 23. NVIDIA also said local LLM downloads are already 5x higher than all of last year.

Key facts

1. Architecture: A 74B-parameter MoE with 9B active parameters per token, squeezed into 24GB via 2-bit quantization with quantization-aware training.
2. Openness: Training data, weights and recipes fully public; launch support for llama.cpp, vLLM, Ollama, Unsloth and LM Studio; direct access via Hermes, OpenClaw and other agent frameworks.
3. Timeline: Bolt 74B ships October 15; RTX Spark laptops go on sale October 16; PAIR leaves beta and goes open source the same day; Bolt E4B 4GB ships October 23.
4. Context: NVIDIA says local LLM downloads are already 5x higher than all of last year; local agent projects such as OpenClaw and Hermes surged in GitHub popularity this year.