Is this page helpful?

Release Notes for NeMo Microservices#

Check out the latest release notes for the NeMo microservices.

Tip

If you’ve installed one of the previous releases of the NeMo microservices using Helm and want to upgrade, choose one of the following options:

Release 25.12.0#

This release includes the following key features and known issues.

Key Features#

Review the key features introduced to the NeMo microservices in this release.

NeMo Studio (Early Access)#

  • Added a Safe Synthesizer interface to create, monitor, and analyze privacy-protected synthetic data generation jobs

  • To learn more about NeMo Studio, refer to the NeMo Studio user guide

NeMo Auditor (Early Access)#

  • Updated the garak security scanner to v0.13.2. This update adds the following new probe.

    • fitd.FITD: Foot in the door probe: Transform queries from benign to harmful obtaining intermediate responses to get compliance from the model.

    The web_injection group of probes are now part of the default/default audit configuration.

    The packagehallucination.PackageHallucinationProbe probe is removed in this release.

    For more information, refer to the probe reference summary.

  • Fixed an issue with the HTML report result. In last release, only the most-recently run probe was shown as the entire probe specification. In this release, all the probes from the audit config are shown in the probe specification.

    Refer to Get an HTML Report for a sample.

NeMo Customizer#

  • Added support for nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16. This is a 30B parameter MoE model (3.5B active) supporting both LoRA and All Weights fine-tuning. See NVIDIA Nemotron 3 Nano 30B A3B for configuration details.

Note

Full SFT Limitation:

  • For All Weights fine-tuning, make sure you configure batch_size == data_parallel_size × micro_batch_size.

  • When deploying the fine-tuned model using the Deployment Management Service, the default NIM profile uses FP8 precision (requires H100+). For A100 GPUs, set NIM_TAGS_SELECTOR: “precision=bf16” in additional_envs to select the BF16 precision profile.

  • Added support for the mistralai/Ministral-3-3B-Reasoning-2512. This model supports both All Weights and LoRA fine-tuning methods.

  • Clarified the batch_size parameter in the documentation, distinguishing it from microbatch_size, and removed the invalid global_batch_size option.

    • Upgraded Automodel to October 2025 version, enabling support for more recent Hugging Face models.

NeMo Data Designer#

  • Support for passing extra params to model providers.

  • Open sourced the core Data Designer library

NeMo Evaluator#

  • Trajectory evaluation of NAT based agents in offline mode

  • Parallelization of LLM-as-a-Judge as a default

  • Usability improvements:

    • Allow job to run and return results even with some NaN outputs

    • Simplify agent evaluation user journey by consolidating evaluation types

    • Updated documentation on secret management and deletion

NeMo Guardrails#

  • Improved code examples throughout the documentation