NVIDIA is releasing large hybrid / MoE models:

- Nemotron 3 Nano: a 30B-parameter model that activates up to 3B parameters per token, aimed at efficient, targeted workloads.

- Nemotron 3 Super: a higher-accuracy reasoning model at ~100B total parameters, with up to 10B active per token, positioned for multi-agent use.

- Nemotron 3 Ultra: a large reasoning model at ~500B total parameters, with up to 50B active per token, aimed at complex AI workflows. Nano is already available:

Super and Ultra are announced for release in the first half of 2026. Announcing that far ahead is a big commitment. They must be pretty confident these will still be competitive when they actually ship.