NVIDIA Nemotron 3 and 3.5: Nano, Super, Ultra and Bolt Are Distinct Releases. This reference distinguishes the original release from hosted endpoints, downloadable weights and subsequent version changes. It is written for people deciding whether an exact model or checkpoint is suitable for a real application, not for leaderboard marketing.
What the sources actually establish
NVIDIA launched the Nemotron 3 family with Nano on December 15, 2025, while presenting Super and Ultra as future releases. Its official research timeline documents Super in March 2026 and Ultra in June. Nemotron 3.5 later expanded the range with Lightning and Bolt variants; the October 15, 2026 announcement described Bolt E4B as becoming available October 23. A 2025 family announcement must therefore not be credited with 2026 checkpoint availability.
A useful application and evaluation scenario
For an internal agent workload, choose an actual named checkpoint and review its required GPUs and tokenizer before calculating deployment capacity. For example, Nemotron 3 Ultra is 550B total and 55B active parameters, whereas the 3.5 Lightning model is a 30B/3B active model. Test retrieval success and tail latency with realistic concurrent requests instead of comparing these numbers as though active parameter count alone controls throughput.
Limits, risks and uncertainties
NVIDIA describes an open-model ecosystem with weights, training assets and recipes, but the precise Nemotron licenses differ by artifact; a separate NIM service is not the same thing as a self-hosted checkpoint. The publisher’s benchmark claims are not independent model tests. Availability of a branded cloud endpoint may differ from direct download or a specific regional enterprise contract.
Alternatives and the decision rule
For edge GPU workloads, start with the documented Bolt sizing; for a large private-cluster reasoning service, test Super or Ultra with infrastructure specialists. Record the exact family-generation and date suffix, because the Nemotron name encompasses small safety models, general language models and multimodal variants that should never share one indiscriminate specification record.
A reproducible test before adoption
Prepare a small held-out set from data the organization is authorized to process. Include simple, ambiguous and deliberately unanswerable requests; specify the expected answer or evidence and a human-reviewed scoring rule. Record model identifier, endpoint and region, tokenizer or SDK version, prompt, tool settings, output budget and whether external retrieval was allowed. Do not present a provider-supplied benchmark as an independently measured result. Separate model mistakes from retrieval, OCR and application validation failures.
Operational checklist and dated evidence
A production transition must test 429 responses, timeouts, refusals, invalid structured output and model-unavailable errors. Keep a reversible rollout or safe fallback instead of silently substituting a model behind an old identifier. Measure latency and usage on the actual workload before planning budgets; join dated token quantities to the AI 10 price ledger. Verify enterprise residency, region, support entitlement and data retention with the actual contract, not a public release story.
Sources & Further Reading
- nvidianews.nvidia.com — documentation/reference (supplier-provided reference; not independently checked in this pass)
- research.nvidia.com — documentation/reference (supplier-provided reference; not independently checked in this pass)
Source documentation establishes vendor-published facts, not independently measured model quality, tenant access, contractual rights or current prices. Use a dated pricing ledger and an account-specific entitlement check for deployment decisions.