NVIDIA released the MCG Toolkit (Model Card Generator) on May 29, according to the NVIDIA developer blog. The tool automates production of model documentation required by the EU AI Act for high-risk AI systems, reducing a process that routinely takes engineering teams days per model to roughly minutes.

The output format is Model Card++, an extended schema that encodes training data lineage, evaluation results, intended use, fairness assessments, and deployment constraints in a single machine-readable document. The toolkit ingests metadata from MLflow, Weights & Biases, and NVIDIA NeMo training workflows, then writes Markdown and JSON variants. Markdown cards publish directly to Hugging Face Hub or the NVIDIA NIM model catalog. The toolkit ships as both a pip-installable library and a NIM microservice, under an Apache 2.0 license.

The compliance angle is the concrete selling point. EU AI Act Annex IV specifies the technical documentation high-risk AI systems must carry before deployment. MCG Toolkit pre-fills those fields from existing training artifacts, giving ML platform teams a defensible paper trail without building custom export pipelines. For enterprises running dozens of production models, that is not a marginal time saving.

The tool lands at an useful moment. EU AI Act obligations for high-risk systems are moving from preparation phase toward enforcement. Compliance teams that have been deferring internal model documentation will find that MCG Toolkit covers the most mechanical part of the requirement: collecting and structuring metadata that already lives in experiment-tracking systems.

NVIDIA’s interest here is not neutral. The Model Card++ format encodes structural choices that favor NVIDIA’s own stack: NeMo is the assumed training lineage, NIM packaging is the deployment target, and the NVIDIA model catalog is the natural registry destination. A documentation standard that, if widely adopted, makes NVIDIA’s infrastructure the path of least resistance is a distribution play as much as a compliance tool. The NVIDIA developer blog does not surface this incentive.

That observation does not diminish the practical value for teams already running NeMo or NIM. For those teams, MCG Toolkit replaces a real chunk of manual documentation work with automated extraction. The cost is committing to NVIDIA’s metadata schema and integration points, which constrains future flexibility if the team moves to a different inference or training stack.

For teams outside the NVIDIA ecosystem, the Model Card++ schema itself is worth examining. If it becomes the dominant format for cross-lab model cards, adapting internal tooling to produce it (regardless of how you generate it) is cheaper than defending a proprietary schema to an auditor.

ML platform teams with EU AI Act obligations in the next twelve months should run the MCG Toolkit against one production model and measure how much of their current Annex IV documentation checklist it covers before deciding whether to adopt it broadly or build against the schema independently.

Source: NVIDIA developer blog, published approximately May 29, 2026.