Skip to main content

Connected AI Platform — Release History

A running, customer-facing index of unified CAIP releases. Entries are grouped by independent offering family and product area, with the newest PI first.

Looking for the full story of one period?

This page gives condensed, area-by-area highlights. Follow the linked full-notes post for the complete customer-facing write-up.

🆕 CAIP Offboard v2: All-New CAIP​

🤖 LLM API​

  • 26-PI3 (May 12 – Jul 6, 2026) — full notes
    • PII and VIN guardrails replace detected sensitive content with placeholders in German and English.
    • Self-hosted chat, embedding, speech-to-text, image-generation, and video-generation models became available.
    • The new /v1/rerank endpoint re-orders retrieved documents by semantic relevance for RAG applications.
    • China LLM API requests now use bearer-token authentication and request-body model selection, with autoscaling, provider fallbacks, and /v1/model/info.
  • 26-PI2 (Mar 17 – May 11, 2026) — full notes
    • The re-architected LLM API entered production with Gateway and OIDC authentication, provider fallbacks, rate limits, and operational dashboards.
    • The catalog added Claude Opus 4.5/4.6, Claude Sonnet 4.6, GPT-5.5, image, video, Bedrock, DeepSeek, Alibaba/Qwen, and Banma models.
    • Integrations now use the updated authorization and model request shape, including UUID-format API keys.
  • 26-PI1 (Jan 20 – Mar 16, 2026) — full notes
    • The catalog added GPT-5 family, GPT Image, Qwen, DeepSeek-V3.2, and image-generation model options.
    • Qwen-Plus and Qwen-Flash now support streaming responses.
  • 25-PI6 (Oct 21, 2025 – Jan 12, 2026) — full notes
    • Translation, GPT-4o-mini speech, and OpenAI-compatible audio transcription became available.
    • Self-hosted Qwen3-8B became available in China, alongside Qwen3 catalog entries.
  • 25-PI5 (Aug 26 – Oct 20, 2025) — full notes
    • LLM API keys became available for customer integrations.
    • Image and embedding models became available in supported environments.
  • 25-PI4 (Jul 1 – Aug 25, 2025) — full notes
    • The LLM API became available in inf-cn and mcaip-inf.
    • OpenAI-compatible embedding requests support the dimensions parameter more reliably.
  • 25-PI3 (May 6 – Jun 30, 2025) — full notes
    • Self-hosted models became available through the LLM API as a platform-managed model option.
  • 25-PI2 (Mar 11 – May 5, 2025) — full notes
    • LLM API authentication moved to the WebEAM realm.
    • API-key rotation became automated.
  • 25-PI1 (Jan 14 – Mar 10, 2025) — full notes
    • LLM API releases 17.0.9, 17.0.10, and 17.1.1 improved stability for existing model integrations.
  • 24-PI6 (Oct 22, 2024 – Jan 13, 2025) — full notes
    • The first LLM API release provided authenticated model access in mcaip-inf through the AI Gateway.
    • Interactive users and machine-to-machine clients could authenticate with OIDC.

🤝 Agents Foundry​

  • 26-PI3 — full notes
    • The dedicated Agents SDK added FAISS vector search for RAG knowledge bases.
    • LangGraph became available for graph-based agent orchestration.
    • Langfuse v4 masks PII in agent traces on both the client and server.
  • 26-PI2 — full notes
    • Agents Foundry launched publicly with the dedicated CAIP Agents SDK and CAIP Agents API.
    • Builders can use PydanticAI or LangChain with conversation threads and streaming.
    • RAG knowledge bases, OpenTelemetry tracing, and Langfuse trace and cost visibility are available.
  • 25-PI6 — full notes
    • The Agents SDK introduced one client for PydanticAI and LangChain.
    • The client supports streaming, conversation threads, and multi-model run and stream operations.

🌐 Web Portal​

  • 26-PI3 — full notes
    • The Model Catalog supports side-by-side comparison and identifies self-hosted models.
    • Doc Assistant added follow-up suggestions, favourites, and lower-latency responses.
    • Space ID validation prevents invalid space configuration at creation time.
  • 26-PI2 — full notes
    • Client-side caching makes the Agents, Cost, and Spaces views more responsive.
    • The Agents list supports server-side pagination, search, sorting, and clearer space navigation.
  • 26-PI1 — full notes
    • The Web Portal added an Admin view.
    • Agent API-key authentication through secrets is simpler, with improved agent-detail navigation and health checks.
  • 25-PI6 — full notes
    • The portal added agent creation, model lookup and listing, and API-key authentication for agent access.
  • 25-PI5 — full notes
    • The portal introduced API-key management, CDH link management, and cost visibility.

📱 CAIP Apps​

  • 26-PI3 — full notes
    • CAIP Apps v1.0.0 is the first stable major release for application deployment on CAIP2.
  • 26-PI2 — full notes
    • CAIP Apps API v0.1.0 is the first public Apps release.
    • It provides endpoints to manage app deployments and trigger image internalization.

⚙️ CAIP Workflows​

  • 26-PI3 — full notes
    • Workflow API releases v0.1 through v0.6 are the first public Workflow API releases.
    • Customers can define stages and assignments, manage configuration, inspect logs, record custom metrics, and validate workflows.
  • 25-PI6 — full notes
    • Workflow definitions can be created, read, updated, and deleted.
    • Run listings, details, logs, workflow API endpoints, and Ray log retrieval became available.

📣 Announcements / Cross-Cutting​

  • 26-PI2 — full notes
    • Connected AI Platform repositories moved to bmw.ghe.com/connected-ai.

🖥️ Classic CAIP (CAIP1 and CAIP1.5)​

🌍 Managed Kubeflow Availability​

  • 25-PI5 (Aug 26 – Oct 20, 2025) — full notes
    • Managed Kubeflow became available in China.
  • 25-PI4 (Jul 1 – Aug 25, 2025) — full notes
    • Managed Kubeflow became available in EMEA.

📓 Notebooks & Development​

  • 26-PI3 (May 12 – Jul 6, 2026) — full notes
    • CAIP SDK 4.8.9–4.8.11 improve GitHub Enterprise and China ECR paths and resolve authentication and pipeline issues.
    • deploy-caip-streamlit-action 2.0.1 supports GitHub Enterprise permissions.
  • 26-PI2 (Mar 17 – May 11, 2026) — full notes
    • Notebook images include MLflow by default for experiment tracking and model-registry work.
    • Code Server and CAIP SDK updates improve GitHub Copilot, credentials, routing, and authentication reliability.
  • 25-PI5 (Aug 26 – Oct 20, 2025) — full notes
    • The CAIP SDK Agent offering added initialization for FastAPI, MCP, and A2A agents, plus a PGVector tool and agent metrics.
    • Teams can validate deployed-agent ingress connections from the CLI and apply stage-specific Helm settings.
    • DNS path-based routes, including MCP subroutes, can expose agents.
  • 25-PI4 (Jul 1 – Aug 25, 2025) — full notes
    • CAIP SDK added agent conversation history, component-store discovery, simpler agent initialization, and per-agent Helm values.
    • Recurring runs can take their parameters from a file.
    • Code Server images included GitHub Copilot extension and chat, and notebook images became available in China.
  • 25-PI3 (May 6 – Jun 30, 2025) — full notes
    • The agent runtime added logging and configuration, stage-scoped environment variables, dependency groups, and deployed-agent listing.
    • caip-runtime supports configurable CDLS-ready logging annotations directly on pipeline steps.
    • enable_do_not_disrupt protects critical workloads, including GPU node-pool workloads, from disruption.
    • TensorFlow and separate ONNX CPU and GPU notebook images became available.
  • 25-PI2 (Mar 11 – May 5, 2025) — full notes
    • CAIP SDK introduced support for deploying and working with agents and added Pydantic v2 compatibility.
    • Artifactory installation and project-packaging helpers became available, replacing Git-token-based setup.
    • Updated notebook images added a newer VS Code Server plus Python 3.10 ONNX and CUDA support.
  • 25-PI1 (Jan 14 – Mar 10, 2025) — full notes
    • CAIP SDK added interactive parameter entry and pipeline selection and refreshed Kubernetes tokens during long-running sessions.
    • Base, SciPy, and Glue notebook images became available with Python 3.12.
  • 24-PI6 (Oct 22, 2024 – Jan 13, 2025) — full notes
    • CAIP SDK added Python 3.10 support for Glue and direct web-interface access to a run using its alias.
    • Cached Docker builds shorten repeat builds.
    • AWS public ECR Python base images avoid Docker Hub rate limits that can delay builds.
  • 24-PI5 (Aug 27 – Oct 21, 2024) — full notes
    • CAIP SDK added multiple parameter sets for comparison runs and arbitrary configuration for each environment.
    • Container builds can run as a non-root user and use dynamic Docker-secret paths.
    • --watch_logs streams run logs to the terminal, and the SDK moved to Python 3.11.
  • 24-PI4 (Jul 2 – Aug 26, 2024) — full notes
    • CAIP SDK added default image tags and shell autocompletion.
    • GitHub Actions can expose direct links to the corresponding Kubeflow pipeline and run.
    • Local pipeline and image versions respond to configuration environment variables.
    • Recurring pipeline schedules can be declared in config.yaml.
  • 24-PI3 (May 7 – Jul 1, 2024) — full notes
    • Python 3.11 notebook images became available for Glue, SciPy, PyTorch, and TensorFlow workloads.
    • CAIP SDK introduced a configuration-driven CLI for multiple pipelines and environments, image build and push, and lifecycle hooks.
    • ECR image builds can use a Dockerfile separate from its build context.
  • 24-PI2 (Mar 12 – May 6, 2024) — full notes
    • Python 3.11 became the default for new CAIP SDK projects.
    • Project artifacts can be uploaded to S3 buckets in custom regions.
    • Python 3.11 Notebook Server and Code Server images became available.
  • 24-PI1 (Jan 16 – Mar 11, 2024) — full notes
    • CAIP SDK added Python 3.8 and 3.11 support, with Python 3.11 the default for new projects.
    • The new MLflow helper marks pipeline steps for experiment tracking.

🔀 Pipelines & Serving​

  • 26-PI3 — full notes
    • The pipeline Teams notifier avoids a webhook error that could prevent notification delivery.
  • 26-PI2 — full notes
    • Glue, EMR, Teams notifier, and run-metrics components were updated for current platform behavior.
    • Ray Service in China supports EFS storage for eligible distributed workloads.
  • 25-PI5 — full notes
    • CAIP SDK restored pipeline updates and secret management, while 3.14.2 restored Kubeflow Pipelines compatibility.
    • Streamlit deployment action 1.0.6 introduced atomic deployments and rollback.
    • Live Inference added model-schema validation, an API client, and a live-prediction endpoint workflow.
  • 25-PI1 — full notes
    • Glue and EMR updates provide clearer job-log context.
    • Streamlit deployment action returns the correct deployment URL in China.
  • 24-PI6 — full notes
    • Glue and EMR components became compatible with CAIP SDK 3.7.1.
    • EMR runs link directly to the Grafana Spark-on-EKS dashboard.
  • 24-PI5 — full notes
    • EMR workflows added retries and environment-based virtual-cluster selection.
    • EMR workflows show the job-run ID and retry count for easier execution tracing.
    • Connected GenAI introduced live inference through Knative and KServe, including a KServe V2 custom-model example.
  • 24-PI4 — full notes
    • A reusable CAIP Streamlit deployment action simplifies consistent application deployment.
    • Run metrics can publish multiple labelled Prometheus metrics alongside run status.
    • EMR workflows added Prometheus metrics and Grafana visualizations for Spark-on-EKS workloads.
  • 24-PI3 — full notes
    • Glue pipeline runs can reuse existing Glue jobs without consuming additional account job capacity.
    • A Spark UI access example and tutorial make Glue Spark jobs easier to inspect.
  • 24-PI2 — full notes
    • Pipeline versioning recognizes relevant YAML changes.
    • Glue components added step retries, Glue 4 by default, and automatic artifact-output paths.
    • EMR components added retries, configurable Spark job overrides, and reliable virtual-cluster selection.
    • A GenAI example environment provides isolated Bedrock model invocation and a KServe custom-model example.
  • 24-PI1 — full notes
    • Container-image builds stream Kaniko logs into component output for visibility in the pipeline UI.
    • Docker-build job failures surface promptly instead of waiting for a timeout.
    • EMR job components link to the correct CloudWatch log location.
    • Foundation models became available through managed Bedrock access, including streamed responses.

🛠 Maintenance​

  • 26-PI3 — full notes
    • The legacy CAIP SDK deployed-agent CLI and API were removed.
    • Customers building AI agents should use CAIP2 Agents Foundry.

📈 Training & Experiment Tracking​

  • 26-PI3 — full notes
    • New guidance explains direct MLflow API access, including authentication and common usage patterns.
  • 25-PI6 — full notes
    • MLflow is installed by default in every Managed Kubeflow namespace.
    • Access reliability fixes and MLflow 3.7 support strengthened experiment tracking and model-registry use.
  • 25-PI5 — full notes
    • MLflow became available through ingress and adopted clearer Managed Kubeflow names.
  • 25-PI3 — full notes
    • MLflow and the model registry became available in additional namespaces.
  • 24-PI3 — full notes
    • Run metrics added named, labelled Gauge and Counter metrics.
  • 24-PI1 — full notes
    • The MLflow helper lets teams mark pipeline steps for experiment tracking.

🗂️ Data & Observability​

  • 26-PI1 — full notes
    • Grafana is enabled by default for Kubeflow profiles.
    • Vector DB became available in more use-case spaces.
  • 25-PI4 — full notes
    • CD4ML buckets use S3 Intelligent-Tiering by default.
    • Vector DB and MLflow became available in additional locations.
  • 25-PI3 — full notes
    • Vector DB became enabled in additional spaces.
    • A supported KubeRay delivery path became available for distributed workloads.
  • 25-PI2 — full notes
    • Managed CDH v2 became available to remaining products and use cases.
  • 24-PI4 — full notes
    • Run metrics can publish multiple labelled Prometheus metrics alongside run status.
    • EMR workflows added Prometheus metrics and Grafana visualizations for Spark-on-EKS workloads.
  • 24-PI3 — full notes
    • Named, labelled pipeline-run metrics became available.
    • A Spark UI access example and tutorial help teams inspect Glue Spark jobs.

🔐 Access & Security​

  • 26-PI3 — full notes
    • The AI-Research to Managed Kubeflow migration is essentially complete, and remaining AI-Research resources are being decommissioned.
  • 26-PI2 — full notes
    • The production Traefik ingress migration is complete.
  • 26-PI1 — full notes
    • Platform configuration requests can be raised through a pull request.
    • The AI-Research to Managed Kubeflow migration program started, with an August 2026 target for remaining AI-Research customers.
  • 24-PI6 — full notes
    • AWS public ECR Python base images avoid Docker Hub rate limits that can delay builds.
  • 24-PI2 — full notes
    • Projects using retired ci-bot credentials receive OIDC migration guidance.
  • 24-PI1 — full notes
    • Pipeline and monitoring components received security updates.

📣 Announcements / Cross-Cutting​

  • 26-PI2 — full notes
    • Connected AI Platform repositories moved to bmw.ghe.com/connected-ai.
  • 26-PI1 — full notes
    • The migration from code.connected.bmw to GitHub Enterprise at bmw.ghe.com/connected-ai was announced.
    • The planned completion target was June 2026.