v2.6.1

This release focuses on model-default documentation and configuration updates. The default NVIDIA-hosted cloud endpoint model changes from Nemotron 3 Super to Nemotron 3 Ultra, and the default self-hosted/on-prem Nemotron 3 Super image is updated to the latest validated 2.0.9 version. This release does not update the RAG Blueprint application container images, Helm chart version, or Python library package version.

Highlights

This release includes the following key updates:

• Default cloud endpoint model changed: NVIDIA-hosted cloud endpoint examples now use nvidia/nemotron-3-ultra-550b-a55b instead of nvidia/nemotron-3-super-120b-a12b. This applies to the LLM, query rewriter, filter expression generator, summarization, reflection, and agentic RAG role examples.
• Default on-prem model updated: Self-hosted/on-prem deployments that use Nemotron 3 Super now reference nvcr.io/nim/nvidia/nemotron-3-super-120b-a12b:2.0.9 instead of 1.8.0. Docker Compose, Helm, MIG, and model-profile examples are updated, including the required NIM_PASSTHROUGH_ARGS=--max-num-seqs 384 Helm override.
• Added Docker Compose environment overlays for Nemotron 3 Ultra local and NVIDIA-hosted deployments.
• Added migration instructions for nvcr.io/nim/nvidia/nemotron-3-embed-1b:2.2.1, including Docker Compose, library mode, hosted endpoint API key configuration, 2048-dimensional embeddings, and re-ingestion guidance.
• Restored Workbench compose image paths to the public nvcr.io/nvidia/blueprint registry organization.

Fixed Known Issues

The following known issues have been resolved in this release:

• Fixed scheduled skills evaluation runs so they execute full sweeps, archive only current-run results, and clean up Brev GPU instances on both success and failure.
• Fixed NVIDIA-hosted rag-blueprint skill evaluation coverage by adding the missing CPU resource metadata.