26-PI3 (May 12 – July 6, 2026) improved the reliability and transition path for CAIP Offboard v1/v1.5 customers while advancing CAIP Offboard v2 model governance, agent development, self-service, Apps, and Workflows. The delivered capabilities make it easier to operate existing workloads securely and build new AI applications on CAIP Offboard v2.

🖥️ CAIP Offboard v1 / v1.5: Classic & Managed Kubeflow​

CAIP Offboard v1: Classic Kubeflow and CAIP Offboard v1.5: Managed Kubeflow customers gained clearer MLflow access, SDK and Streamlit deployment improvements, and an orderly transition away from legacy deployed-agent tooling.

✨ New Features​

📓 Notebooks & Development​

CAIP SDK releases 4.8.9–4.8.11 improve GitHub Enterprise and China ECR paths, resolve authentication and pipeline issues, and include security updates. deploy-caip-streamlit-action 2.0.1 supports GitHub Enterprise permissions, helping teams continue to deploy Streamlit applications under the current permission model.

📈 Training & Experiment Tracking​

New guidance explains direct MLflow API access for Managed CAIP users, including authentication and common usage patterns. Teams can use the API more confidently when integrating experiment tracking and model-registry workflows with their own tools.

🔐 Access & Security​

The AI-Research to Managed Kubeflow migration is essentially complete, and the remaining AI-Research resources are being decommissioned. This marks the transition of customer workloads to the managed platform.

🛠 Maintenance​

The legacy deployed-agent CLI and API in the CAIP SDK were removed. Customers building AI agents should use Agents Foundry, CAIP Offboard v2's dedicated agent-development offering, as their successor path. The pipeline Teams notifier also avoids a webhook error that could prevent notifications from being delivered.

📦 Repositories in This Release​

• caip-sdk — SDK reliability, security, and legacy deployed-agent removal.
• deploy-caip-streamlit-action — GitHub Enterprise permission support.
• kubeflow-components — notifier delivery reliability.

This section covers customer-visible changes for CAIP Offboard v1/v1.5 from May 12 – July 6, 2026 (26-PI3). For guides and reference material, see the CAIP Documentation.

🆕 CAIP Offboard v2: All-New CAIP​

CAIP Offboard v2 customers gained production guardrails and a wider self-hosted model portfolio, more capable Agents Foundry tools, stronger portal discovery, and the first public Workflow API releases.

✨ New Features​

🤖 LLM API​

Production PII and VIN guardrails now replace detected sensitive content with placeholders in German and English, helping customers protect inputs and outputs while preserving request flow. The LLM API also added self-hosted chat, embedding, speech-to-text, image-generation, and video-generation models, plus Qwen3.6-35B-A3B, additional DashScope image models, and self-hosted text-to-speech. The new /v1/rerank endpoint lets RAG applications re-order retrieved documents by semantic relevance.

The China LLM API re-architecture standardizes bearer-token authentication and model selection in the request body while retaining the validity of existing China API keys. Its catalog added DeepSeek, Qwen, and image models, and retired qwen2.5-32b-instruct, qwen3-4b, deepseek-v3, and wan2.6-t2i. Autoscaling, provider fallbacks, expanded observability, and /v1/model/info improve dependable model selection and catalog discovery across regions.

🤝 Agents Foundry​

The dedicated Agents SDK added FAISS vector search for RAG knowledge bases and LangGraph for graph-based agent orchestration. Langfuse v4 now masks PII in agent traces on both the client and server, giving teams richer observability while helping protect sensitive data.

🌐 CAIP Web Portal​

The Model Catalog supports side-by-side comparison and labels self-hosted models, making it easier to choose an appropriate option. The Doc Assistant adds follow-up suggestions, favourites, and lower-latency responses, while Space ID validation prevents invalid space configuration at creation time.

📱 CAIP Apps​

CAIP Apps v1.0.0 is the first stable major Apps release, providing customers with a stable supported foundation for application deployment on CAIP Offboard v2.

⚙️ CAIP Workflows​

The Workflow API had its first public releases, v0.1 through v0.6. Customers can define workflow stages and assignments, manage configuration, inspect logs, record custom metrics, and validate workflows, giving them an API-led way to operate and understand workflow execution.

🛠 Maintenance​

Portal-service authorization now uses a fail-closed, OPA-based model, strengthening access control when authorization decisions cannot be safely completed.

📦 Repositories in This Release​

• caip-llm-api — guardrails, self-hosted models, reranking, and regional reliability.
• caip-agents-sdk and caip-langfuse — agent orchestration, vector search, and trace masking.
• caip-web-portal, caip-apps-api, and caip-workflow-api — portal, Apps, and Workflow API releases.

This section covers customer-visible changes for CAIP Offboard v2 from May 12 – July 6, 2026 (26-PI3). For guides and reference material, see the CAIP Documentation.

26-PI2 (March 17 – May 11, 2026) strengthened the daily development experience in CAIP Offboard v1/v1.5 and brought major new CAIP Offboard v2 services into production. The PI combines more reliable notebooks and pipelines with a re-architected LLM API, the first public Agents Foundry and Apps releases, and a more capable portal.

🖥️ CAIP Offboard v1 / v1.5: Classic & Managed Kubeflow​

CAIP Offboard v1: Classic Kubeflow and CAIP Offboard v1.5: Managed Kubeflow customers gained a ready-to-use MLflow experience, more reliable development tooling, updated pipeline integrations, and the completed production ingress migration.

✨ New Features​

📓 Notebooks & Development​

Notebook images now include MLflow by default, so experiment tracking and model-registry work can start without an additional installation step. Code Server updates improve GitHub Copilot conversation reliability, while CAIP SDK fixes address AWS credentials, ingress and Traefik base URLs, and authentication refreshes—making notebook and SDK workflows more dependable across environments.

🔀 Pipelines & Serving​

Updated Glue, EMR, Teams notifier, and run metrics components help teams run and observe established pipeline integrations with current platform behavior. Ray Service in China now supports EFS storage, giving eligible distributed workloads shared persistent storage.

🔐 Access & Security​

The production Traefik ingress migration is complete, providing the supported ingress path for customer access. Connected AI Platform repositories have also moved to bmw.ghe.com/connected-ai; customers and contributors should use the new GitHub Enterprise location for repositories and pull requests.

📦 Repositories in This Release​

• caip-sdk and jupyter-notebooks-images — notebook, credential, routing, and authentication reliability.
• kubeflow-components and products — pipeline integrations, EFS support, and ingress migration.

This section covers customer-visible changes for CAIP Offboard v1/v1.5 from March 17 – May 11, 2026 (26-PI2). For guides and reference material, see the CAIP Documentation.

🆕 CAIP Offboard v2: All-New CAIP​

CAIP Offboard v2 introduced production-ready foundations for model access, agent development, and application deployment, while improving the portal experience used to manage them.

✨ New Features​

🤖 LLM API​

The re-architected LLM API entered production with Gateway and OIDC authentication, automatic provider fallbacks, rate limits, and operational dashboards. Its catalog now includes Claude Opus 4.5/4.6, Claude Sonnet 4.6, GPT-5.5, gpt-image-1.5, gpt-image-2, sora-2, and additional Bedrock, DeepSeek, Alibaba/Qwen, and Banma models, giving applications broader language and multimodal options. Integrations use the updated authorization and model request shape, including UUID-format API keys, to access the production API consistently.

🤝 Agents Foundry​

Agents Foundry had its first public launch with the dedicated CAIP Agents SDK and CAIP Agents API. Builders can use PydanticAI or LangChain with conversation threads, streaming, RAG knowledge bases, OpenTelemetry tracing, and Langfuse visibility for traces and costs. SDK v1.0.1 also resolves early integration issues, making the new agent-development path more reliable from the start.

🌐 CAIP Web Portal​

Client-side caching makes the Agents, Cost, and Spaces views more responsive. The portal was updated from Angular 19 to 21, and the Agents list now supports server-side pagination, search, sorting, and clearer space navigation, helping customers find and manage resources as their estates grow.

📱 CAIP Apps​

CAIP Apps API v0.1.0 is the first public Apps release. It provides endpoints to manage app deployments and trigger image internalization, giving application teams a supported API foundation for bringing their workloads onto CAIP Offboard v2.

📦 Repositories in This Release​

• caip-llm-api — production API rollout, catalog expansion, resilience, and observability.
• caip-agents-sdk and caip-agents-api — Agents Foundry launch.
• caip-web-portal — faster resource management and Agents-list improvements.
• caip-apps-api — Apps API v0.1.0.

This section covers customer-visible changes for CAIP Offboard v2 from March 17 – May 11, 2026 (26-PI2). For guides and reference material, see the CAIP Documentation.

26-PI1 (January 20 – March 16, 2026) made CAIP Offboard v1/v1.5 easier to observe and manage while expanding CAIP Offboard v2's LLM API choices and portal administration. Customers can use more of the platform with less setup, from Kubeflow dashboards to multimodal model access.

🖥️ CAIP Offboard v1 / v1.5: Classic & Managed Kubeflow​

CAIP Offboard v1: Classic Kubeflow and CAIP Offboard v1.5: Managed Kubeflow customers gained more immediate visibility, broader data-service availability, and a clearer way to request platform configuration. This PI also marked the start of the AI-Research to Managed Kubeflow migration program and the communication of the upcoming GitHub move.

✨ New Features​

📓 Notebooks & Observability​

Grafana is now enabled by default for Kubeflow profiles. Teams can open dashboards for their notebooks and workloads without first arranging a separate observability setup, making it easier to monitor day-to-day platform use.

🗂️ Data​

Vector DB is available in more use-case spaces, extending semantic-search and retrieval-augmented workloads to additional teams. This gives customers more places to build applications that combine their data with similarity search.

🔐 Access & Platform Evolution​

Platform configuration requests can now be raised directly through a pull request. The change provides a clearer, reviewable path for requesting configuration changes and keeps the request together with the proposed configuration.

The AI-Research to Managed Kubeflow migration program started this PI, with an August 2026 target for remaining AI-Research customers. We also announced the migration of Connected AI Platform repositories from code.connected.bmw to GitHub Enterprise at bmw.ghe.com/connected-ai, planned for June 2026. These two transitions establish the supported platform and collaboration paths going forward.

📦 Repositories in This Release​

• products — default Grafana, Vector DB availability, and configuration-request improvements.

This section covers customer-visible changes for CAIP Offboard v1/v1.5 from January 20 – March 16, 2026 (26-PI1). For guides and reference material, see the CAIP Documentation.

🆕 CAIP Offboard v2: All-New CAIP​

CAIP Offboard v2 customers can select from a substantially broader LLM catalog and manage agents more easily in the Web Portal.

✨ New Features​

🤖 LLM API​

The LLM API catalog added the GPT-5 family—GPT-5, GPT-5 mini, GPT-5 nano, GPT-5.2, GPT-5 Codex, GPT-5 Pro, o3, and o3 Pro—alongside GPT Image 1 and GPT Image 1 mini. New Qwen3, Qwen3.5, DeepSeek-V3.2, Qwen Image, Qwen Image Max, Qwen Image Plus, and Wan2.6-T2I options give builders more choices for chat, reasoning, and image-generation use cases. Qwen-Plus and Qwen-Flash now also support streaming, so applications can present responses as they are generated.

🌐 CAIP Web Portal​

The Web Portal added an Admin view and simplified agent API-key authentication through secrets, reducing the effort to administer agent access. Corrected breadcrumb navigation on agent details and improved health checks make the portal easier to navigate and more dependable in daily use.

🛠 Maintenance​

Azure OpenAI API keys were rotated across environments as part of routine credential hygiene, preserving secure access to the service.

📦 Repositories in This Release​

• products — LLM API catalog and credential updates.
• caip-web-portal — administration, agent authentication, navigation, and health improvements.

This section covers customer-visible changes for CAIP Offboard v2 from January 20 – March 16, 2026 (26-PI1). For guides and reference material, see the CAIP Documentation.

25-PI6 (October 21, 2025 – January 12, 2026) made MLflow a default part of Managed Kubeflow and expanded CAIP Offboard v2 with richer capabilities for agents, workflows, portal access, and multimodal LLM use.

🖥️ CAIP Offboard v1 / v1.5: Classic & Managed Kubeflow​

✨ New Features​

📈 Training & Experiment Tracking​

MLflow became installed by default in every Managed Kubeflow namespace, making experiment tracking and the model registry immediately available across the managed experience. Access reliability fixes and MLflow 3.7 support further strengthened customers' ability to use these capabilities.

📦 Repositories in This Release​

• products — default MLflow installation and access reliability.

This section covers customer-visible changes for CAIP Offboard v1/v1.5 from October 21, 2025 – January 12, 2026 (25-PI6). For guides and reference material, see the CAIP Documentation.

🆕 CAIP Offboard v2: All-New CAIP​

✨ New Features​

🤝 Agents Foundry​

The Agents SDK introduced one client for PydanticAI and LangChain, with streaming, conversation threads, and multi-model run and stream operations. Customers could use these capabilities to work with supported agent frameworks through a more unified client experience.

🌐 Web Portal​

The portal added agent creation, model lookup and listing, and API-key authentication for agent access, giving customers a direct place to discover models and manage how agents connect.

⚙️ Workflows​

Workflow definition creation, reading, updating, and deletion became available alongside run listings, details, logs, workflow API endpoints, and Ray log retrieval. This gave customers a fuller set of workflow controls and the operational details needed to follow their runs.

🤖 LLM API​

The LLM API added translation, GPT-4o-mini speech, and OpenAI-compatible audio transcriptions, broadening the supported multimodal use cases. Self-hosted Qwen3-8B became available in China, and Qwen3 catalog entries helped customers discover the models available to them.

📦 Repositories in This Release​

• products — portal, workflow, and LLM API capabilities.
• caip-agents-sdk — dedicated agent client capabilities.

This section covers customer-visible changes for CAIP Offboard v2 from October 21, 2025 – January 12, 2026 (25-PI6). For guides and reference material, see the CAIP Documentation.

What’s Changing?​

• AWS CLI and BMWAWS CLI login are deprecated.
• Use the orbit-use CLI for all future logins to AWS accounts and CAIP clusters.

How to Login​

To access CAIP, you must assume the CAIP developer role for your teamspace.

Role:

<orbit-space>-<orbit-teamspace>-caip-developer

Use the following command:

orbit-use cloud <orbit-space> <dev|prod> <orbit-teamspace> --role caip-developer

Examples:

orbit-use cloud smartdaip dev genai-tsara --role caip-developer

orbit-use cloud ai-rsrch dev aira --role caip-developer

To open the Orbit web interface after login, add the --ui option:

orbit-use cloud <orbit-space> <dev|prod> <orbit-teamspace> --role caip-developer --ui

Examples:

orbit-use cloud smartdaip dev genai-tsara --role caip-developer --ui

orbit-use cloud ai-rsrch dev aira --role caip-developer --ui

If only one role is available in the teamspace, you can omit --role and orbit-use will auto-detect it:

orbit-use cloud <orbit-space> <dev|prod> <orbit-teamspace>

Why This Matters​

The new CLI streamlines authentication, improves security, and integrates seamlessly with cloud CLI tools. Profiles are automatically configured for you, making it easier to manage credentials and access.

More Information​

For full details and step-by-step instructions, see our updated IAM documentation.

If you have any questions or need support, please refer to our support documentation.

Thank you for adapting to this change and helping us keep our platform secure and efficient!

We are announcing the removal of Amazon Bedrock access from the CAIP Managed Platform. As part of this change, the Bedrock:InvokeModel action will be removed from all policies within our scope.

The execution date for this change will be announced soon.

What is changing?​

• Bedrock:InvokeModel permissions will be revoked from all managed IAM policies.
• Customers will no longer be able to invoke Bedrock models directly through the CAIP Managed Platform.
• Customers are advised to migrate their workloads to Connected AI Platform's LLM API.

Why are we making this change?​

This update is part of our ongoing efforts to streamline and secure our platform, ensuring that all customers use a consistent and supported interface for large language model (LLM) access.

What should you do?​

If you are currently using Bedrock through the CAIP Managed Platform, please migrate to our LLM API. Our LLM API provides a robust, secure, and fully supported way to access large language models.

• LLM API Documentation
• LLM API Model Catalogue

If you have any questions or need assistance with the migration, please contact our support team.

Thank you for your understanding and cooperation as we continue to improve our platform.

What is the LLM API?​

The LLM API is the Connected AI Platform’s unified interface for accessing state-of-the-art generative AI models. It enables customers to use chat completion, embeddings, and image generation models seamlessly in both Rest of World (RoW) and China regions.

With the LLM API, you can:

• Access several Amazon Bedrock models (e.g., Claude, Nova, Llama)
• Use models from Azure OpenAI Service (e.g., GPT-4o, GPT-3.5)
• Leverage Alibaba Cloud models (e.g., Qwen, DeepSeek)
• Integrate with self-hosted models

This unified API simplifies integration, provides consistent authentication, and supports both production and China-specific endpoints. For more details, see the LLM API Documentation and the LLM API Model Catalogue.

25-PI5 (August 26 – October 20, 2025) launched Managed Kubeflow in China and advanced CAIP Offboard v1/v1.5 pipeline, experiment-tracking, CAIP SDK agent, and Live Inference capabilities, while CAIP Offboard v2 added customer-facing API and portal controls.

🖥️ CAIP Offboard v1 / v1.5: Classic & Managed Kubeflow​

✨ New Features​

🌍 Managed Kubeflow Availability​

Managed Kubeflow is now available in China, giving teams in the region a managed, supported foundation for developing, running, and operating ML workloads.

🔀 Pipelines & Serving​

CAIP SDK 4.6.0 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, so unsuccessful deployments did not leave an application partially updated.

🤖 CAIP SDK Agents​

The CAIP SDK Agent offering added initialization for FastAPI, MCP, and A2A agents, a PGVector tool, MCP integration improvements, and agent metrics. Teams could also validate a deployed agent's ingress connection directly from the CLI, apply stage-specific Helm settings for different environments, and expose agents through DNS path-based routes, including MCP subroutes.

🎯 Live Inference​

Live Inference added model-schema validation, an API client, and a live-prediction endpoint workflow. These capabilities made it easier for CAIP Offboard v1/v1.5 teams to validate a model integration and connect an application to a deployed live-inference service.

📈 Training & Experiment Tracking​

MLflow became available through ingress and adopted clearer Managed Kubeflow names, making the experiment-tracking service easier for customers to find and access.

📦 Repositories in This Release​

• caip-sdk — pipeline updates, secrets, Kubeflow Pipelines compatibility, and CAIP SDK Agent capabilities.
• deploy-caip-streamlit-action — atomic Streamlit deployments and rollback.
• products — Managed Kubeflow China availability and MLflow access improvements.
• caip-inference-api — model validation, live-inference client, and prediction endpoint workflow.

This section covers customer-visible changes for CAIP Offboard v1/v1.5 from August 26 – October 20, 2025 (25-PI5). For guides and reference material, see the CAIP Documentation.

🆕 CAIP Offboard v2: All-New CAIP​

✨ New Features​

🤖 LLM API​

LLM API keys became available, giving customers a dedicated way to authenticate their LLM API integrations. Image and embedding models also became available in the supported environments, expanding the model capabilities teams could use through the platform.

🌐 Web Portal​

The portal introduced API-key management, CDH link management, and cost visibility, bringing these important customer controls and information into one experience.

📦 Repositories in This Release​

• products — LLM API key enablement and image/embedding model availability.
• caip-portal-api and caip-web-portal — API-key management, CDH links, and cost visibility.

This section covers customer-visible changes for CAIP Offboard v2 from August 26 – October 20, 2025 (25-PI5). For guides and reference material, see the CAIP Documentation.

25-PI4 (July 1 – August 25, 2025) launched Managed Kubeflow in EMEA, improved CAIP Offboard v1/v1.5 agent development and notebook access, and extended LLM API availability and compatibility for more customer use cases.

🖥️ CAIP Offboard v1 / v1.5: Classic & Managed Kubeflow​

✨ New Features​

🌍 Managed Kubeflow Availability​

Managed Kubeflow is now available in EMEA, giving teams in the region a managed, supported foundation for developing, running, and operating ML workloads.

📓 Notebooks & Development​

CAIP SDK 4.2.0–4.5.1 added agent conversation history, component-store discovery and documentation, simpler agent initialization, protected intranet exposure, and per-agent Helm values. Recurring runs could also take their parameters from a file, making scheduled executions easier to repeat consistently without manually re-entering settings. Code Server notebook images included the GitHub Copilot extension and chat, and notebook images were published in China, expanding access to the available development environment.

🗂️ Data & Observability​

CD4ML buckets began using S3 Intelligent-Tiering by default, while Vector DB and MLflow became available in additional locations, extending these managed data and experiment-tracking capabilities.

📦 Repositories in This Release​

• caip-sdk — agent history, discovery, initialization, deployment configuration, and parameter-file-driven recurring runs.
• jupyter-notebooks-images — Code Server and China image publication.
• products — Managed Kubeflow EMEA availability, Intelligent-Tiering, Vector DB, and MLflow availability.

This section covers customer-visible changes for CAIP Offboard v1/v1.5 from July 1 – August 25, 2025 (25-PI4). For guides and reference material, see the CAIP Documentation.

🆕 CAIP Offboard v2: All-New CAIP​

✨ New Features​

🤖 LLM API​

The LLM API became available in inf-cn and mcaip-inf, expanding where customers could use it. OpenAI-compatible embedding requests also supported the dimensions parameter more reliably, improving compatibility for those integrations.

📦 Repositories in This Release​

• products — LLM API availability and embedding compatibility.

This section covers customer-visible changes for CAIP Offboard v2 from July 1 – August 25, 2025 (25-PI4). For guides and reference material, see the LLM API Documentation.

Help Us Make Connected AI Platform Even Better!

Your opinion matters to us! We are constantly working to improve Connected AI Platform and make it more useful, reliable, and enjoyable for you. That’s why we’d love to hear from you.

Please take just 5 minutes of your time to share your thoughts and suggestions.
Your feedback is voluntary, anonymous, and will directly help shape the future of our application.

📌 Your input makes a real difference – thank you for supporting us!
⏳ Survey available until: 30.09.2025.

ℹ️ We do not store any personal information about you. We use a standardized feedback service provided by UX DesignOps (UX Live Center), which focuses on general usability and does not allow detailed, application-specific questions.

Deprecate code servers with development environment with python 3.8 version. 

Release Highlights​

It will not be possible to use python v3.8 due to BMW politics. Now only it is possible to create code server notebooks with python v3.11 and python v.12.

How to Upgrade​

The notebooks with the server code with python v3.8 will stop working. You will need to delete the notebooks to not use unnecessary resources and then create new code servers with the new versions (v3.11 or v3.12).

Please note:

• To ensure a smooth transition, we give users time to upgrade to more updated versions.
• The version will be deprecated in one week, 2025.7.31.

The goal of this change is, mainly, to cut costs across all of the cd4ml buckets, since it's been noticed that we have a lot of stored objects, sometimes summing up to terabytes of used space.

Release Highlights​

All cd4ml buckets will transition it's current storage (S3 Standard) to S3 Intelligent-Tiering. This feature is turned on by default, as of the date of this change. Settings can be defined accordingly in the product.yaml, following the structure below.

The currently existing objects in these buckets will not be deleted, but transferred to another storage tiering. This is done via a lifecycle policy that defines a 0 day transition to intelligent-tiering, as shown below.

This change aims to try to cut costs in the long run by automatically transferring unaccessed data to a lower-cost tier. The defined settings for unaccessed data are the following:

• 0 - 30 days: Frequent Access Tier
• 30 - 90 days: Infrequent Access Tier
• 90 - 120 days: Archive Instant Access Tier
• 120 - 150 days: Archive Access Tier
• 150+ days: Deep-Archive Access Tier

Everytime data is accessed in any of these tiers, it returns back to the 'Frequent Access Tier'.

⚠️ Main Takeaways and Important Notes ⚠️​

While we advise maintaining data in Intelligent-Tiering, here are some takeaways that may be important to mention: 

• These changes are applied to all cd4ml buckets. For now, there is no option to "cherry-pick" buckets to not use it;
• Objects under 128 kb are always kept in the 'Frequent Access Tier';
• Instant Access Tiers are currently opt-out via product.yaml. If you turn this feature OFF:
  • All objects stored in intelligent tiering won't be moved do S3 Standard / other unless a manual copy is made;
  • New objects inserted after instant access is turned off will be stored in S3 Standard;
  • If you turn this feature ON again, objects in Standard will be moved back to Intelligent Tiering automatically during the next evaluation cycle;
• Archive Access Tiers are currently opt-in via product.yaml. If you turn these features ON:
  • Data that is in 'Archive Instant Acess', if it's unaccessed time is above defined thresholds (120+ / 150+), will be moved to their respective tiers during the next evaluation cycle;
  • Data retrieval from Archive Access Tiers require a restore request, which causes it to be brought back to 'Frequent Access Tier';
  • If you turn these features OFF again, objects will stay in Intelligent Tiering, in their respective tiers. Retrieval of this data to another storage type will require a restore and manual copy;

We also advise to revise bucket storage usage to try to reduce storage costs whenever possible (temp files, unused data / files, etc...).

25-PI3 (May 6 – June 30, 2025) broadened CAIP Offboard v1/v1.5 agent and notebook capabilities, expanded data and experiment-tracking availability, and added self-hosted models as a managed LLM API option.

🖥️ CAIP Offboard v1 / v1.5: Classic & Managed Kubeflow​

✨ New Features​

📓 Notebooks & Development​

CAIP SDK 4.1.0–4.1.5 extended the agent runtime with logging and configuration, stage-scoped environment variables, dependency groups, and a way to list deployed agents. The new caip-runtime also let teams add configurable CDLS-ready logging annotations directly to pipeline steps, while enable_do_not_disrupt protected critical workloads—including GPU node-pool workloads—from disruption. Notebook images added TensorFlow and separate ONNX CPU and GPU builds, giving teams more targeted runtime choices.

🗂️ Data & Observability​

Vector DB became enabled in additional spaces, while MLflow and the model registry reached more namespaces, bringing data and experiment-tracking capabilities to more teams. A supported KubeRay delivery path also became available for distributed workloads.

📦 Repositories in This Release​

• caip-sdk — agent runtime configuration, pipeline logging annotations, workload protection, and discovery.
• jupyter-notebooks-images — TensorFlow and ONNX images.
• products — Vector DB, MLflow, model-registry, and KubeRay availability.

This section covers customer-visible changes for CAIP Offboard v1/v1.5 from May 6 – June 30, 2025 (25-PI3). For guides and reference material, see the CAIP Documentation.

🆕 CAIP Offboard v2: All-New CAIP​

✨ New Features​

🤖 LLM API​

Self-hosted models became available through the LLM API, giving customers an additional platform-managed model option alongside the API access they already used.

📦 Repositories in This Release​

• products — self-hosted models through the LLM API.

This section covers customer-visible changes for CAIP Offboard v2 from May 6 – June 30, 2025 (25-PI3). For guides and reference material, see the LLM API Documentation.

25-PI2 (March 11 – May 5, 2025) expanded CAIP Offboard v1/v1.5 agent development, notebook options, and managed data connectivity, while making LLM API access management easier to operate.

🖥️ CAIP Offboard v1 / v1.5: Classic & Managed Kubeflow​

✨ New Features​

📓 Notebooks & Development​

CAIP SDK 4.0.0–4.0.3 introduced native support for deploying and working with agents, Pydantic v2 compatibility, and helpers for Artifactory installation and project packaging. The SDK moved installation away from Git-token-based setup and clarified its CLI/API command namespaces, making upgrades and day-to-day use easier to navigate. Updated notebook images added a newer VS Code Server plus Python 3.10 ONNX and CUDA support, broadening the tools available for development.

🗂️ Data & Observability​

Managed CDH v2 became available to the remaining products and use cases, extending consistent, managed data connectivity across CAIP for more customer scenarios.

📦 Repositories in This Release​

• caip-sdk — agent support, Pydantic v2 compatibility, Artifactory installation, clearer command namespaces, and packaging helpers.
• jupyter-notebooks-images — VS Code Server, ONNX, and CUDA updates.
• products — Managed CDH v2 availability.

This section covers customer-visible changes for CAIP Offboard v1/v1.5 from March 11 – May 5, 2025 (25-PI2). For guides and reference material, see the CAIP Documentation.

🆕 CAIP Offboard v2: All-New CAIP​

✨ New Features​

🤖 LLM API​

The LLM API authentication realm moved to WebEAM, and API-key rotation became automated. Together, these changes kept customer credentials current without requiring manual rotation work.

📦 Repositories in This Release​

• products — WebEAM authentication and API-key rotation.

This section covers customer-visible changes for CAIP Offboard v2 from March 11 – May 5, 2025 (25-PI2). For guides and reference material, see the LLM API Documentation.

25-PI1 (January 14 – March 10, 2025) made CAIP Offboard v1/v1.5 pipeline development and job investigation more approachable, refreshed notebook images, and strengthened reliability for existing LLM API integrations.

🖥️ CAIP Offboard v1 / v1.5: Classic & Managed Kubeflow​

✨ New Features​

📓 Notebooks & Development​

CAIP SDK releases 3.11.0–3.14.0 resolved local image-build issues, added interactive parameter entry and pipeline selection, and refreshed Kubernetes tokens automatically during long-running sessions. Base, SciPy, and Glue notebook images also became available with Python 3.12, giving teams current options for common development workloads.

🔀 Pipelines & Serving​

Glue 2.3.2 and EMR 1.8.1 delivered clearer job-log context, helping customers investigate submitted jobs with more useful information at hand. Streamlit deployment action 1.0.4 also returned the correct deployment URL in China, keeping the deployment handoff dependable.

📦 Repositories in This Release​

• caip-sdk — local builds, interactive pipeline use, and token refresh.
• jupyter-notebooks-images — Python 3.12 base, SciPy, and Glue images.
• kubeflow-components — Glue and EMR job-log context.
• deploy-caip-streamlit-action — corrected China deployment URLs.

This section covers customer-visible changes for CAIP Offboard v1/v1.5 from January 14 – March 10, 2025 (25-PI1). For guides and reference material, see the CAIP Documentation.

🆕 CAIP Offboard v2: All-New CAIP​

✨ New Features​

🤖 LLM API​

LLM API releases 17.0.9, 17.0.10, and 17.1.1 improved stability for existing model integrations, reinforcing a more dependable experience for customers already using the API.

📦 Repositories in This Release​

• products — LLM API stability releases.

This section covers customer-visible changes for CAIP Offboard v2 from January 14 – March 10, 2025 (25-PI1). For guides and reference material, see the LLM API Documentation.

24-PI6 (October 22, 2024 – January 13, 2025) improved CAIP build and scheduling reliability, extended Glue support, and marked CAIP Offboard v2's first LLM API release with authenticated model access.

🖥️ CAIP Offboard v1 / v1.5: Classic & Managed Kubeflow​

✨ New Features​

📓 Notebooks & Development​

The CAIP SDK added Python 3.10 support for Glue and let teams open a run directly in the web interface with a run alias. Cached Docker builds shortened repeat builds, while AWS public ECR Python base images avoided Docker Hub rate limits that could otherwise delay them.

🔀 Pipelines & Serving​

Glue and EMR components became compatible with CAIP SDK 3.7.1. EMR runs also linked directly to the Grafana Spark-on-EKS dashboard, helping teams move from a pipeline run to the operational information relevant to that workload.

🛠 Maintenance​

Cron scheduling validation was corrected, helping scheduled pipelines reject invalid expressions before they run.

📦 Repositories in This Release​

• caip-sdk — Glue Python 3.10 support, web run access, cached builds, scheduling validation, and public ECR base images.
• kubeflow-components — Glue and EMR compatibility plus Grafana dashboard links.

This section covers customer-visible changes for CAIP Offboard v1/v1.5 from October 22, 2024 – January 13, 2025 (24-PI6). For guides and reference material, see the CAIP Documentation.

🆕 CAIP Offboard v2: All-New CAIP​

✨ New Features​

🤖 LLM API​

The LLM API had its first release, delivering authenticated model access in mcaip-inf through the AI Gateway. Both interactive users and machine-to-machine clients could authenticate with OIDC to use the available models, establishing a managed entry point to this new CAIP Offboard v2 capability.

📦 Repositories in This Release​

• products — first LLM API release with authenticated AI Gateway access.

This section covers customer-visible changes for CAIP Offboard v2 from October 22, 2024 – January 13, 2025 (24-PI6). For guides and reference material, see the LLM API Documentation.

24-PI5 (August 27 – October 21, 2024) advanced comparison workflows, build reliability, and live inference in CAIP Offboard v1/v1.5, giving teams more confidence and visibility across the delivery journey.

🖥️ CAIP Offboard v1 / v1.5: Classic & Managed Kubeflow​

✨ New Features​

📓 Notebooks & Development​

The CAIP SDK added multiple parameter sets for comparison runs and arbitrary configuration for each environment, so teams could evaluate alternatives without duplicating projects. Container builds could run as a non-root user and use dynamic Docker-secret paths, while --watch_logs streamed run logs directly to the terminal for immediate visibility. The SDK also moved to Python 3.11, keeping project development on a supported, maintained runtime.

🔀 Pipelines & Serving​

EMR workflows added retries and environment-based virtual-cluster selection, then surfaced the EMR job-run ID and retry count so teams could trace a run's execution state more easily. Connected GenAI also introduced live inference through Knative and KServe, with a KServe V2 example to help teams integrate a deployed custom model.

📦 Repositories in This Release​

• caip-sdk — comparison runs, environment configuration, resilient image builds, log streaming, and Python 3.11 support.
• kubeflow-components — EMR retries, environment-aware virtual-cluster selection, and execution details.
• products — Connected GenAI live inference and KServe integration example.

This section covers customer-visible changes for CAIP Offboard v1/v1.5 from August 27 – October 21, 2024 (24-PI5). For guides and reference material, see the CAIP Documentation.

24-PI4 (July 2 – August 26, 2024) streamlined CAIP project delivery and brought richer operational metrics to pipeline workloads, helping teams move from delivery to insight with less friction.

🖥️ CAIP Offboard v1 / v1.5: Classic & Managed Kubeflow​

✨ New Features​

📓 Notebooks & Development​

The CAIP SDK added default image tags and shell autocompletion to reduce friction when building and running projects. GitHub Actions could expose direct links to the corresponding Kubeflow pipeline and run, while local pipeline and image versions responded to configuration environment variables—making project activity easier to connect with its pipeline context. Recurring pipeline schedules could now be declared in config.yaml, keeping execution timing version-controlled, reviewable, and reproducible alongside the rest of a project configuration.

🔀 Pipelines & Serving​

A reusable CAIP Streamlit deployment action made consistent Streamlit application deployment simpler. The run-metrics component could publish multiple labelled Prometheus metrics alongside run status, and EMR workflows added Prometheus metrics with Grafana visualizations for Spark-on-EKS workloads, providing a more complete operational view.

📦 Repositories in This Release​

• caip-sdk — image tags, autocompletion, workflow links, configuration-aware versions, and YAML-defined schedules.
• deploy-caip-streamlit-action — reusable Streamlit deployment action.
• kubeflow-components — labelled metrics, run status, and EMR observability.

This section covers customer-visible changes for CAIP Offboard v1/v1.5 from July 2 – August 26, 2024 (24-PI4). For guides and reference material, see the CAIP Documentation.

24-PI3 (May 7 – July 1, 2024) made CAIP projects more configurable, broadened Python 3.11 notebook support, and gave teams clearer ways to inspect Glue workloads and pipeline metrics.

🖥️ CAIP Offboard v1 / v1.5: Classic & Managed Kubeflow​

✨ New Features​

📓 Notebooks & Development​

Python 3.11 notebook images became available for Glue and SciPy workloads as well as PyTorch and TensorFlow development, giving teams a broader choice of current runtimes. The CAIP SDK introduced a configuration-driven CLI for multiple pipelines and environments, image build and push, lifecycle hooks, proxy configuration, and environment-variable placeholders. Teams could also use a Dockerfile that sits separately from its build context for ECR image builds, avoiding unnecessary project restructuring. Richer configuration objects made those settings easier to reuse across project code.

🔀 Pipelines & Serving​

Glue pipeline runs gained the ability to reuse existing Glue jobs, helping teams continue running workflows without consuming additional account job capacity. A Spark UI access example and tutorial also made Glue Spark jobs easier to inspect when teams needed operational detail.

📈 Training & Experiment Tracking​

The run-metrics component added named, labelled Gauge and Counter metrics, enabling teams to publish the measurements that matter most to their experiments and workflows.

📦 Repositories in This Release​

• jupyter-notebooks-images — Python 3.11 Glue, SciPy, PyTorch, and TensorFlow images.
• caip-sdk — configuration-driven CLI, flexible ECR image builds, and project configuration support.
• kubeflow-components — Glue job reuse, Spark UI guidance, and run metrics.

This section covers customer-visible changes for CAIP Offboard v1/v1.5 from May 7 – July 1, 2024 (24-PI3). For guides and reference material, see the CAIP Documentation.

24-PI2 (March 12 – May 6, 2024) strengthened the reliability and flexibility of CAIP workflows, expanded modern notebook choices, and introduced a focused environment for exploring GenAI capabilities.

🖥️ CAIP Offboard v1 / v1.5: Classic & Managed Kubeflow​

✨ New Features​

📓 Notebooks & Development​

The CAIP SDK made Python 3.11 the default for new projects and added artifact uploads to S3 buckets in custom regions, giving teams more flexibility in where project outputs are stored. New Python 3.11 Notebook Server and Code Server images also brought a current runtime to interactive development.

🔀 Pipelines & Serving​

Pipeline versioning began recognizing relevant YAML changes, helping teams run the intended version when a definition 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 selection of the requested virtual cluster—giving workflow authors more resilient, configurable jobs.

🤖 Foundation Models​

A GenAI example environment introduced isolated access for Bedrock model invocation, giving teams a focused place to explore the capability. Its KServe custom-model example also reported readiness after a model loaded, making it clear when the example could receive requests.

🛠 Maintenance​

The SDK warns projects still using retired ci-bot credentials to move to OIDC, supporting a safer authentication path. Local-version and run-completion handling were also made more resilient.

📦 Repositories in This Release​

• caip-sdk — Python 3.11 defaults, regional S3 uploads, pipeline-version handling, and OIDC migration guidance.
• jupyter-notebooks-images — Python 3.11 Notebook Server and Code Server images.
• kubeflow-components — Glue and EMR workflow improvements.
• products — GenAI example environment and KServe custom-model readiness.

This section covers customer-visible changes for CAIP Offboard v1/v1.5 from March 12 – May 6, 2024 (24-PI2). For guides and reference material, see the CAIP Documentation.

24-PI1 (January 16 – March 11, 2024) gave CAIP teams a stronger foundation for building, observing, and operating pipelines, while opening managed access to foundation models for eligible users.

🖥️ CAIP Offboard v1 / v1.5: Classic & Managed Kubeflow​

✨ New Features​

📓 Notebooks & Development​

The CAIP SDK added Python 3.8 and 3.11 support, with Python 3.11 becoming the default for new projects. Teams could also use the new MLflow helper to mark pipeline steps for experiment tracking more directly, bringing experiment visibility closer to the workflow definition.

🔀 Pipelines & Serving​

Kubeflow Pipelines made build and job investigation more transparent: container-image builds streamed Kaniko logs into component output, so teams could follow progress in the pipeline UI, while EMR job components linked to the correct CloudWatch log location for faster failure investigation. Docker-build job failures also surfaced promptly instead of leaving teams waiting for a timeout, so broken builds were quicker to diagnose and fix.

🤖 Foundation Models​

Foundation models became available through the platform's managed Bedrock access, including streamed responses. Improved model discovery and entitlement handling helped eligible teams identify the models available to them and put that managed access to work.

🛠 Maintenance​

Pipeline and monitoring components received security updates, helping keep common workflow building blocks current and secure.

📦 Repositories in This Release​

• caip-sdk — Python 3.8/3.11 support and the MLflow step helper.
• kubeflow-components — build-log streaming, corrected EMR log links, and component security updates.
• products — managed foundation-model access and streamed responses.
• terraform-modules — model discovery, availability, and entitlement support.

This section covers customer-visible changes for CAIP Offboard v1/v1.5 from January 16 – March 11, 2024 (24-PI1). For guides and reference material, see the CAIP Documentation.