NVIDIA spent the first half of 2026 rolling out Nemotron 3, an open-weights model family built on a hybrid Mamba-Transformer mixture-of-experts architecture and aimed at long-running agentic workloads. The line staged out from Nemotron 3 Super (120B total parameters, 12B active) in March to Nemotron 3 Ultra (550B total, 55B active) in June, with a multimodal Nano Omni in between. NVIDIA released not just the weights but the training data and recipes, positioning Nemotron 3 as a genuinely open frontier-scale family rather than a demo.

What happened

The releases came in sequence. Nemotron 3 Super arrived on 11 March 2026, featured at GTC 2026, with a native 1-million-token context window and pretraining largely in NVFP4 4-bit precision for Blackwell hardware. Nemotron 3 Nano Omni (30B total, 3B active) followed on 28 April 2026 as a single open model unifying video, audio, image, and text. Nemotron 3 Ultra, NVIDIA’s largest open-weights model to date, shipped on 4 June 2026 under the permissive Linux Foundation OpenMDW-1.1 license.

At GTC on 16 March 2026 NVIDIA also launched the Nemotron Coalition, a collaboration with eight AI labs (including Mistral AI, Perplexity, Cursor, and Thinking Machines) to co-train an open base model on DGX Cloud, positioned as the foundation for a future Nemotron 4. The hybrid Mamba-Transformer design is the technical thread: it mixes state-space and attention layers to cut the cost of very long contexts.

Why it matters for builders

Open weights plus open data plus a documented recipe is a different offer from a downloadable checkpoint. It lets teams fine-tune, audit, and self-host frontier-scale models without a vendor API in the loop, which matters for regulated, sovereign, or cost-sensitive deployments. The Mamba-Transformer architecture is also a signal: pure-attention transformers are no longer the only frontier design, and the hybrid approach targets exactly the long-context agentic workloads that are getting expensive to serve.

If you are choosing an open model to build on, Nemotron 3 now spans sizes from a 30B multimodal Nano to a 550B Ultra, so you can match the tier to your hardware. Pair this with what inference costs and the 2026 LLM landscape to weigh it against other open-weight families.

Sources

Further reading