Aleph Alpha

Provides LLM inference (Luminous family models) executed entirely on EU-hosted infrastructure with transparent data handling policies and GDPR compliance built into the platform architecture. Requests never leave European data centers, and data retention policies are explicitly configurable per deployment. The infrastructure implements strict data isolation at the hypervisor level and provides audit logs for regulatory compliance verification.

Built-in capability to visualize which input tokens influenced each output token through attention weight extraction and attribution analysis. The platform exposes attention maps from the Luminous model's transformer layers, allowing developers to trace decision paths and understand model reasoning at the token level. This is implemented as a first-class API feature, not a post-hoc analysis tool, enabling real-time explainability in production systems.

Luminous models support extended context windows (up to 2048 tokens for base models, 4096+ for extended variants) enabling processing of longer documents and conversations. The platform provides utilities for managing context, including automatic summarization of long conversations, sliding window techniques for maintaining context across multiple turns, and efficient token counting to avoid exceeding context limits.

Enables organizations to fine-tune Luminous base models on proprietary datasets to adapt the model for domain-specific tasks (e.g., legal document analysis, medical terminology) while maintaining data privacy. Fine-tuning is performed on customer infrastructure or Aleph Alpha's EU-hosted environment with full data isolation. The platform provides managed fine-tuning pipelines with hyperparameter optimization, validation set handling, and version control for model checkpoints.

Provides tools and APIs for systematically engineering prompts and few-shot examples to improve model performance on specific tasks. The platform includes prompt templating, example management, and A/B testing capabilities to compare prompt variants. Developers can structure examples with explicit input/output formatting, and the API supports dynamic prompt construction based on retrieval or user context.

Enables semantic search over document collections using Aleph Alpha's embedding models, which rank documents by semantic similarity rather than keyword matching. The platform provides APIs to embed documents, store embeddings, and retrieve top-k results for a given query. Embeddings are generated using the same Luminous architecture as the language models, ensuring semantic consistency across the platform.

Supports processing of documents beyond plain text, including PDFs, images, and structured data formats. The platform can extract text from documents, understand layout and structure, and pass document content to language models for analysis. This enables use cases like document classification, information extraction from forms, and visual question answering on document images.

Provides configurable safety filters and content moderation capabilities that can be tuned to organizational policies. The platform allows teams to define custom guardrails (e.g., blocking specific topics, enforcing tone constraints) and apply them to model outputs. Safety filtering is transparent and explainable — the system indicates which guardrail was triggered and why, rather than silently filtering content.

Supports batch processing of multiple requests in a single API call, with asynchronous execution and cost discounts for non-real-time workloads. Developers can submit batches of prompts, receive a job ID, and poll for results. Batch processing is optimized for throughput rather than latency, enabling cost-effective processing of large document collections or bulk analysis tasks.

Provides granular rate limiting, quota management, and transparent pricing per API call. Organizations can set rate limits per API key, monitor usage in real-time, and receive alerts when approaching quota limits. Pricing is transparent and usage-based — no surprise charges or hidden fees. The platform provides detailed cost breakdowns per request type and model variant.

Luminous models support multiple languages (English, German, French, Spanish, and others) with language-specific variants optimized for each language. The platform automatically detects input language and routes to the appropriate model variant. Language-specific models are trained on language-native data, improving performance on non-English tasks compared to English-only models.