An open reasoning model with its recipe beside it.
NVIDIA released Nemotron 3 Super, an open 120-billion-parameter mixture-of-experts reasoning model with 12 billion active parameters, plus stated releases of its methodology, data, reinforcement-learning environments, and evaluation recipes.
- Maturity
- Deployed, stage 4 of 4
- Support
- 1 source · institution
- Evidence detail
- How we know ↓
Long-running, multi-step AI work can become slow and expensive when each agent repeatedly carries a large task history and uses a large model for every subtask.
Open model releases often provide weights without enough data and training information for researchers to inspect or reproduce how a system was made.
A hybrid Mamba-transformer mixture-of-experts model for complex agent subtasks, with a one-million-token context window and multiple active expert specialists at inference.
What was shown, and what wasn’t
Shown
NVIDIA says the model was available on March 11, 2026 with open weights under a permissive license. It says it published methodology, more than 10 trillion pre- and post-training tokens, 15 reinforcement-learning training environments, and evaluation recipes. NVIDIA reports up to 5 times higher throughput and up to 2 times higher accuracy than the prior Nemotron Super model.
Not shown · limits
- The throughput, accuracy, and agent-quality results are NVIDIA's claims, not independent evidence of benefit in a workplace or public service.
- The underlying training data includes synthetic data from frontier reasoning models; openness does not itself settle provenance, bias, or safety questions.
- A one-million-token context window can retain more material, but it does not prevent mistakes, goal drift, or unsafe tool use.
- Hardware requirements and operational cost can still limit who can use or customize the model.
Bright editorial interpretation
The combination of weights, large stated data release, training environments, and recipes could let more researchers inspect, reproduce, and adapt a modern reasoning-model pipeline rather than only consume a hosted tool.
Still open
Can independent groups reproduce the claimed throughput and accuracy gains?
How complete and usable are the published data and recipes in practice?
Do long-context agents improve real work quality, safety, and accountability rather than simply produce longer traces?
How we know1 source · checked 2026-09-07 · no corrections
Original sources
Institutions: NVIDIA
- Maturity
- Deployed
- Source published
- 2026-03-11
- Captured
- 2026-09-07
- Last source review
- 2026-09-07
- Editorial method
- AI-assisted source review
Bright compared this account with the linked original and supporting sources and kept reported, budgeted, projected, and observed claims distinct. Bright did not independently audit the underlying records.
Maturity describes the tested or operational setting. Confidence describes support for the particular claim; one does not determine the other.
Revision & correction history
2026-09-07 · The combination of weights, large stated data release, training environments, and recipes could let more researchers inspect, reproduce, and adapt a modern reasoning-model pipeline rather than only consume a hosted tool.
No corrections recorded.
Keep exploring
Explore the shared question in another setting. These connections do not imply replication.