NVIDIA's Nemotron 3 Puzzle Runs Big on a Lean Budget

NVIDIA has published Nemotron-Labs-3-Puzzle-75B-A9B, a new open-weight reasoning model that leans on a sparse mixture-of-experts design to keep inference costs down. The model carries 75 billion total parameters but activates only about 9 billion per token, according to its Hugging Face repository. The weights are distributed in BF16.

Why it matters

The appeal of a latent-MoE layout is straightforward: you get the knowledge capacity of a large model without paying to run all of it at once. That trade-off has become a common thread across recent open releases, and Nemotron 3 Puzzle slots into that trend by targeting reasoning workloads specifically, where a slimmer active footprint can translate into faster, cheaper chains of thought.

A few things stand out from the release:

• Total vs. active split: 75B parameters on disk, roughly 9B engaged per token.
• Reasoning focus: the model is tagged for text generation and reasoning tasks.
• BF16 weights: shipped in a format suited to modern accelerators.

NVIDIA has been steadily expanding its Nemotron line, and this "Puzzle" variant signals continued investment in sparse architectures aimed at reasoning rather than raw scale. The company has not published a context length for this checkpoint, and the license is listed as a custom NVIDIA agreement, so teams should read the terms before building on it.

For developers weighing open models, the pitch here is efficiency: a large knowledge base that behaves, at runtime, like something much smaller. Full details and the model card are available on Hugging Face.