🚀 What’s New after 25-PI4?
The last two months have been an exciting period for the Connected AI Platform. With every PI we’re moving closer to our vision of a unified, reliable, and developer-friendly AI platform at BMW. In this update, we want to share with you the highlights of what we shipped in 25-PI4 — from new portal features to powerful APIs, and a glimpse of what’s coming next.
✨ New Features
🌐 CAIP Web Portal
We are excited to announce the first release of the CAIP Web Portal — a completely new entry point to the Connected AI Platform. The portal will serve as the foundation for bringing all CAIP offerings together in one place.
With this first version, you can already use:
- Space Management: A completely new concept that allows you to bundle upcoming CAIP features into logical units — making it easier to organize and manage your projects in one place.
- User Management: Full control over who can collaborate in your space.
- Managed Kubeflow Integration: Direct access to CAIP’s managed Kubeflow service from within the Web Portal.
This is just the beginning — more features will be added step by step in upcoming releases.
⚙️ Managed Kubeflow
That means you can now run Kubeflow on CAIP’s managed platform without worrying about setup, scaling, or maintenance.
With Managed Kubeflow you can:
- Start with experiments and prototyping.
- Scale up seamlessly as your workloads grow.
- Stay on the same managed environment all the way into production.
This closes the gap between research and production, giving you a reliable path from idea to deployment.
We are proud to introduce Managed Kubeflow. It is fully operated by the CAIP platform team and designed to support both experimentation and production workloads in a secure, scalable way. Instead of setting up and maintaining your own cluster, you can now rely on a platform-managed, production-grade environment from day one.
🔑 Key Benefits of Managed Kubeflow:
- Fully integrated into the CAIP Self-Service Web Portal.
- Fast, reliable onboarding through service requests.
- Namespace-based isolation instead of per-cluster setups.
- Production infrastructure only — no more separate INT/E2E clusters.
- Built-in staging concept for pipelines and models (within one infra).
- Optimized for performance, observability, and security.
- Backed by standard CAIP SLAs.
- Seamless go-live without migration between environments.
👉 If you’re starting a new ML use case today, you should use Managed Kubeflow.
🔗 New CDH Integration
As the Connected AI Platform (CAIP) transitions towards a managed architecture, the integration with the Cloud Data Hub (CDH) is evolving as well.
The new approach is built on two key concepts:
- Role chaining
- CDH-managed environments
This setup makes the interaction between CAIP and CDH datasets more scalable and secure, while reducing the operational burden on all involved teams.
By shifting responsibility for dataset environments to CDH and relying on role chaining for access, we:
- Simplify the process of consuming CDH datasets from within CAIP.
- Reduce complexity in managing CDH resources directly on CAIP.
- Prepare the IAM layer for the upcoming migration to a Managed CAIP architecture.
In short: this integration not only improves today’s workflows, but also lays the foundation for the next generation of CAIP.
🤖 LLM API (Beta)
The new LLM API is live in Beta and offers powerful large language model capabilities. It currently provides two main endpoints: Chat Completions and Embeddings.
💬 Chat Completions
A dedicated Chat Completions endpoint is now available via /v1/chat/completions.
This endpoint allows you to access Large Language Models (LLMs) — via a unified interface, across multiple model providers.
Supported Model Families
-
Rest of World (RoW)
- Claude (Anthropic) — reasoning-focused conversational models.
- GPT (OpenAI via Azure) — GPT-3.5 and GPT-4.x families, including lightweight variants.
- Nova (AWS Bedrock) — optimized for speed, cost, and scalability.
- Qwen (Ali Cloud) — Turbo and Long variants for extended context.
- DeepSeek — R1 and V3 families for performance and efficiency.
- LLaMA (Meta) — lightweight and open models at different parameter scales.
-
China
- Qwen (Ali Cloud) — Qwen+, Turbo, Long, and 72B families.
- DeepSeek — R1, V3, and distilled models.
🧭 Embeddings
The LLM API now also provides an Embeddings endpoint via /v1/embeddings.
Embeddings convert text into high-dimensional vectors that capture semantic meaning — enabling use cases like semantic search, clustering, classification, or retrieval-augmented generation (RAG).
Supported Model Families
-
Rest of World (RoW)
- Azure OpenAI — text-embedding-3 family (small & large).
-
China
- Ali Cloud — text-embedding-v4.
👉 You can find the full and always up-to-date model catalog in our documentation.
📣 Communication & Enablement
We’ve also invested in how we share knowledge, updates, and processes — making CAIP easier to discover and navigate.
🔑 Key Updates
- Documentation Hub — A centralized entry point to all CAIP docs, offering consistent navigation and improved search.
- CAIP Update Blog — Regular updates, deep dives, and best practices directly from the platform team.
- New Confluence Home — A redesigned landing page for better structure and easier access to resources.
- New Processes — Clear and accountable handling of incidents (INC), service requests (SR), with SLAs. L2 enablement for faster resolution and support.
💡 Feature Requests: Powered by You
Your feedback is essential for shaping the Connected AI Platform. In this PI, we’ve picked up and delivered some of the most requested improvements from the community.
🔑 Highlights
- Bigger Notebooks — Support for notebooks with up to 64 GB RAM (CONNECTAI-6937).
- Smarter Storage — Intelligent Tiering for CD4ML S3 Buckets, optimizing cost and performance (CONNECTAI-6943).
- GitHub Copilot — Available on VS Code Servers to boost developer productivity (CONNECTAI-6966).
🛠 Maintenance Behind the Scenes
Not all improvements are visible at first glance — but they make a big difference in your daily work.
In this PI we focused on strengthening the platform foundation with a series of performance and reliability updates.
🔑 Key Improvements
- Performance Enhancements — Faster startup times and smoother execution for notebooks and pipelines.
- Reliability Gains — Reduced failure rates and more predictable behavior under heavy workloads.
- Operational Stability — Improved monitoring and alerting to detect issues earlier and recover faster.
🛠 Maintenance
While new features grab the headlines, we also spent a lot of time improving the foundation. In this PI we focused on:
- Performance improvements across notebooks and pipelines.
- Reliability enhancements, ensuring smoother day-to-day operation.
These may not always be visible at first glance, but they’re crucial to keep Connected AI Platform stable and efficient as usage grows.
🔭 Outlook: What’s Cooking?
We’re already working on the next wave of CAIP capabilities — designed to make the platform even more powerful, production-ready, and easy to use.
-
🚀 CAIP Workflows: A new way to orchestrate and automate ML workflows directly in CAIP.
-
⚡ Distributed AI Compute (via CAIP Workflows): Scale AI workloads efficiently across clusters with a dedicated Workflow API.
-
🧩 Managed Inference API: Easily deploy models into production with a fully managed inference layer.
-
📦 CAIP Apps: Pre-packaged applications that extend the CAIP ecosystem with ready-to-use solutions.
-
🗂️ CAIP Feature Store: Centralized management and reuse of ML features across projects, improving consistency and speed.
-
🔑 Portal Improvements: Upcoming enhancements to the CAIP Web Portal, including long-awaited API Key Management for easier integration and automation.
]]>
