AI Model Hub

What Is It?

AI Model Hub is a fully-managed inference service that provides on-demand access to pre-trained large language models (LLMs), embedding models, rerankers, text-to-image generators, and OCR models via OpenAI-compatible REST APIs. The service eliminates the need to provision GPUs, manage scaling, or maintain model servers. All processing occurs exclusively in IONOS German data centers (Berlin) with GDPR compliance, ISO 27001 certification, and data sovereignty. Customer data is never used for model training.

Quick Facts

Active Model Catalog

Retired Models (do not use)

Model Capabilities

Embedding Models

What You Can Do

Text Generation

Run pre-trained LLMs for conversational AI, Q&A systems, content creation, and chatbots. All text generation models support streaming responses for real-time interactions.

OCR (Optical Character Recognition)

Extract text from images using LightOnOCR-2-1B. The model accepts image input (PNG, JPEG) and returns Markdown-formatted text including LaTeX for math. PDFs are not directly supported; render each page to an image before calling the API. A pdf-to-text how-to guide is available in the documentation.

Note: LightOnOCR-2-1B outputs Markdown format only; this cannot be changed via the API.

Image Generation

Generate images from textual prompts using FLUX.1-schnell via /v1/images/generations. Supported sizes: 1024x1024, 1024x1792, 1792x1024. FLUX.2-klein 4B additionally supports image editing via /v1/images/edits and is billed per megapixel.

Tool Calling

Enable models to invoke external APIs or predefined functions for dynamic automation. Use for workflow triggers, real-time data retrieval, or business application integration. Available on most LLMs (Mistral Nemo, Mistral Small 24B, Llama 3.1 8B, Llama 3.3 70B, Llama 3.1 405B, GPT-OSS 120B, Qwen3 Coder Next 80B).

Semantic Embeddings and Reranking

Convert text or images into dense vector representations using BGE Large v1.5, BGE M3, MPNet v2, or Qwen3-VL-Embedding-8B. Use for similarity search, clustering, duplicate detection, or recommendation engines. Reranking (Qwen3-VL-Reranker-8B) re-scores retrieved passages before generation to improve RAG precision.

Retrieval-Augmented Generation (RAG)

Build RAG pipelines by combining Hub embedding and LLM models with an external vector database. The recommended path for new builds is IONOS Managed PostgreSQL with the pgvector extension.

Note: The native Document Collections feature (managed Chroma DB vector store) is deprecated and closed to new deployments (end of life 2026-08-31). Existing collections continue to function until that date, after which they are no longer available via the API; new RAG projects should use the external pgvector path.

Best For

Consider Alternatives If

Key Considerations

Billing & Costs

• LLM billing: Per million input + output tokens (see Active Model Catalog table for per-model rates)
• OCR billing: EUR 0.15/M input tokens, EUR 0.30/M output tokens (LightOnOCR-2-1B)
• Embedding billing: EUR 0.01-0.10/M tokens depending on model (see Embedding Models table)
• Reranker billing: EUR 0.04/M tokens (Qwen3-VL-Reranker-8B)
• Image generation: EUR 0.0288/image (FLUX.1-schnell)
• Document collection storage (legacy): EUR 0.01 per million tokens per 30 days
• Cost optimization: Use smaller models (8B-12B) for latency-sensitive or high-throughput workloads

Limitations

• Stateless service: Prompts and outputs are discarded after each request; no conversation history stored server-side
• Document Collections deprecated: Closed to new deployments (end of life 2026-08-31); use Managed PostgreSQL + pgvector for new RAG builds
• Document input format (legacy collections): Plain text only; PDFs and Word documents must have text extracted before upload; max 65,535 characters per document
• OCR input: LightOnOCR-2-1B accepts images only; PDFs must be converted to images page-by-page first
• Image generation rate limit: FLUX.1-schnell is capped at 10 images/minute (burst to 20)
• No cross-region replication: All processing fixed to Germany; not suitable where data must stay outside Germany
• Authentication: Bearer Token required; generate via DCD > Management > Token Manager.

Compliance

• GDPR: Input data never reused for model training; service logs kept only for operational health
• EU AI Act: IONOS acts as Distributor for unmodified open-source models, and as AI Provider (with additional transparency obligations) for quantized models such as Llama 3.1 405B-FP8
• Customer obligations: Limited-risk applications must display an AI interaction notice; high-risk applications require customer-implemented logging, human-in-the-loop, and data governance

Management Options

• OpenAI-compatible clients: Python openai package, curl, or any HTTP client
• Native REST API: Model management, legacy document collection CRUD, and RAG-specific endpoints
• Authentication: API tokens via DCD > Management > Token Manager.
• Supported integrations: Any language or framework that can make HTTP REST calls