Can you own it?
This page is a projection of the one entry record, the Use & modify and Transparency factors that Assess covers. The full verdict is set by all four factors together, floor-weighted so the weakest caps the whole.
- AssessUse & modify + Transparency
- ImplementData control + Reliability
- UseReliability
- SupportTransparency
Intended & out-of-scope use
Nemotron 3 Ultra is the frontier-scale head of the family: 550B total / 55B active, a Mamba2-Transformer hybrid Latent Mixture-of-Experts with Multi-Token Prediction, a context window up to 1M tokens. Its intended use is complex agentic workflows, long-context analysis and high-end reasoning where you want an open, maximally-permissively-licensed model and can run a multi-node cluster.
Out of scope: EU high-stakes deployment until the systemic-risk position is resolved. Training compute is undisclosed, so whether Ultra crosses the 1e25-FLOP threshold cannot be established. The licence (OpenMDW-1.1) grants use "without restriction", so the gating questions are hardware, safety assurance and systemic-risk status.
Known limitations, bias & failure modes
The released Ultra recipe is "a representative single pass". Its intermediate checkpoints "have not been open-sourced" and the 1M-context phase is excluded because its data "is not open-source", so the model is not reproducible. There is no dedicated model-level safety evaluation: the White Paper's "Evaluation, Safety and Release" section is a contributor list, and safety appears only as post-training data curation. Only "approximately 20 trillion tokens" is stated for training compute, so the EU systemic-risk crossing is undeterminable. At 550B / 55B active there is no small-variant portability, and Ultra is a multi-GPU / multi-node deployment.
The offsetting advantages are the most permissive licence in the family and open-weights-recipe transparency.
Openness tier & components
Ultra is open_weights_recipe (tier 4). Weights are ungated safetensors, training code is a
representative recipe (Pretrain -> SFT -> MOPD -> Quant), the documentation includes a white paper,
and post-training data is CC-BY-4.0. It stops short of fully_open because the recipe is explicitly
representative: the intermediate checkpoints it depends on are not open-sourced, and the 1M-context
data is not open, so the model is not fully reproducible.
License terms & permitted use
Ultra is governed by the OpenMDW License Agreement v1.1 (OpenMDW-1.1), authored by the OpenMDW project under the Linux Foundation and adopted by NVIDIA. It is the most permissive of the three Nemotron 3 licences: the grant is "permission is hereby granted, free of charge, to deal in the Model Materials without restriction", covering weights, data, documentation and software together. Commercial use is allowed with no field-of-use limit. The only condition is a defensive termination that fires if the licensee brings, maintains or voluntarily joins a patent or copyright infringement lawsuit over the Model Materials, with an explicit carve-out where that suit answered one brought first against them. It is not OSI-approved because v1.1 is not on the OSI approved list. This "without restriction" grant is what makes Ultra's use-and-modify and data-control factors strong.
The raw licence file has no acceptable-use, guardrail or revocation clause. OpenMDW-1.1 was submitted for OSI review in August 2026 and no approval is listed, so it is recorded as not OSI-approved.
Supply-chain & provenance
Weights are distributed from the verified nvidia org on Hugging Face as ungated safetensors,
with official BF16 / NVFP4 quantization variants and a clear canonical source. The checkpoint trust
checklist scores about 4/8: published checksums, checkpoint scanning and a signing/attestation
policy are not verified, which is why provenance is a 3. Pin the exact revision and verify
checksums.
EU AI Act posture
Ultra is GPAI. Its OpenMDW licence is maximally permissive, but the EU information base is the same as the rest of the family: no formal GPAI documentation package, no copyright policy, and NVIDIA is not a Code of Practice signatory. The open issue is systemic risk: training compute is undisclosed, so whether a 550B / 55B-active model trained on ~20T tokens crosses the 1e25-FLOP threshold is undeterminable. An EU deployer should treat systemic-risk status as unresolved before high-stakes use.
Benchmarks & evaluation
Ultra is a frontier-scale open model (550B / 55B active, hybrid Mamba2-Transformer LatentMoE, up to 1M context, ~20T tokens). OneHill has not run its own benchmarks, and no independent third-party re-runs exist, so the figures are publisher and architectural (white paper + card). This is marked partial and holds the performance dimension at 3.
Independent safety evaluation
There is no dedicated model-level safety evaluation for Ultra: the White Paper's "Evaluation, Safety and Release" section is a contributor list, and safety otherwise appears only as post-training data curation (Nemotron Content Safety v2, Gretel refusal data, keyword/regex filtering). The downloadable family guard stack - the Nemotron-3-Content-Safety classifier, Apache-2.0 NeMo Guardrails, and Apache-2.0 garak - is attributable to Ultra and holds the safety dimension at 3, not 2. The missing model-level evaluation holds it below 4. The guard-classifier scores are NVIDIA self-reported.
Ultra has a dedicated technical report (arXiv 2606.15007, June 2026). Its safety content is
training-side: a safety SFT blend translated into six languages and an "agentic safety teacher"
trained against indirect prompt injection with an automated red-teaming loop. It publishes no
benchmark results, jailbreak results or residual-risk disclosure, and the safety.md subcard gives
no scores either. There is no published model-level safety evaluation.
How this scores
The ownership factors this domain covers, drawn from the one entry record.
Use and modify freelyCan you run, modify and adapt it with no gate and no field-of-use trap?
StrongUltra is under OpenMDW-1.1, whose grant is to 'deal in the Model Materials without restriction' - an MIT/BSD-style, unconditional permission covering weights, data, documentation and software, with commercial use allowed and no acceptable-use or field-of-use limit. Weights are ungated and fine-tunable. The only condition is a defensive litigation termination (with a first-sued carve-out), which does not restrict ordinary use or modification. This is the permissive-open case, and the reason Ultra is the family member that reaches strong here.
TransparencyDo you know what it is: weights, training, behaviour, and legible terms?
Strongopen_weights_recipe: ungated weights, a white paper and a representative training recipe with CC-BY post-training data let you see a great deal of how Ultra was built. The intermediate checkpoints and the 1M-context data are not open, which limits reproducibility and caps the openness tier. Visibility into the build is high.
Sources
The same evidence records as the entry sheet. Read means the text was verified; unverified means it is known to exist but not yet read.