NVIDIA Puts a 30B Agentic Model on a Single Laptop GPU

NVIDIA released Nemotron 3.5 Lightning and NeMo Switchyard on 11 August 2026, extending the Nemotron 3 family that already includes the Nemotron 3 Ultra model on our roster.

Nemotron 3.5 Lightning

A 30-billion-parameter mixture-of-experts model with roughly 3 billion parameters active per token. NVIDIA positions it as the highest-efficiency model in its class for long-running agentic workloads and claims up to four times the output speed of comparable models. It is open source and runs on a single GPU - including in a laptop or desktop, not just a datacentre.

At 3B active out of 30B total, inference cost tracks a small model while capacity tracks a much larger one, which is what makes single-GPU agentic loops practical.

NeMo Switchyard

An open-source routing library that directs each step of an agent workflow to the most capable and efficient model available for that step.

Nemotron 4 is in training

NVIDIA is training a next generation whose largest member is expected to reach at least one trillion parameters. There is no announced release date; reporting at the time suggested it could be ready as early as late autumn 2026. Treat that as unconfirmed - NVIDIA has not committed to a date.