Kolibri reasoning and tool calling
Kolibri is built for assistants and agentic workflows that need multi-step reasoning, structured extraction, retrieval-augmented generation (RAG), and reliable tool calling.
Aleph Alpha builds specialized language models and enterprise AI software for public-service, document, and engineering workflows. Teams can run its platform as SaaS or on their own infrastructure.
Aleph Alpha develops specialized language models and enterprise AI software for public institutions and industry. Its current model, Kolibri, supports multi-step reasoning, structured extraction, retrieval-augmented generation (RAG), and tool calling in German and English. The offer also includes a platform, ready-made AI workspaces, and an engineering workbench.
The platform connects document collections and existing systems for searchable, source-cited answers, and can run as sovereign SaaS, in a hybrid setup, on premises, or air-gapped. Workspace Essentials combines enterprise chat with agents, skills, and connectors; the Industrial Engineering Workbench compares requirements documents and shows cited suggestions that engineers can approve or correct. Kolibri's downloadable weights are licensed under Apache 2.0; this model artifact is separate from the enterprise platform and inference offering.
Kolibri is built for assistants and agentic workflows that need multi-step reasoning, structured extraction, retrieval-augmented generation (RAG), and reliable tool calling.
Document AI extracts files, tables, charts, technical drawings, and scans into usable data. Teams can search connected content with access controls and get answers that quote source passages; the system can abstain when that content is insufficient to answer.
The workspace combines enterprise chat with agents, skills, and connectors that work in existing systems. Each answer cites its sources and is traceable to a named user, model, and prompt.
The Industrial Engineering Workbench analyzes, structures, and compares requirements specifications in large documents. Engineers can inspect cited suggestions and approve or correct them in the application; it connects to PLM or runs as a Siemens Polarion add-on.
The Adaptive Platform serves, routes, and scales models across sovereign cloud, hybrid, and on-premises environments. It also offers Responses API inference and state management.
Use Kolibri in agent workflows for tasks such as retrieving a record, checking eligibility criteria, or filling in a form through connected tools and systems.
Make files, tables, charts, technical drawings, and scans searchable, then answer questions from an organization’s own information with cited source passages.
Review requirements across large document sets and use source-linked suggestions to see what has already been addressed, then approve or correct proposals in the workbench.
Teams can run the platform as sovereign SaaS, in a hybrid setup, on premises, or air-gapped, depending on their deployment needs.
Kolibri is designed for assistants and agentic workflows using multi-step reasoning, structured extraction, retrieval-augmented generation, and tool calling. For example, a citizen-service agent can retrieve a record, check eligibility criteria, or fill in a form.
Kolibri supports German and English and is specialized in German.
Yes. The Adaptive Platform can run as sovereign SaaS, in a hybrid setup, on premises, air-gapped, or fully self-managed.
Kolibri’s FP8 model weights have a footprint of about 78 GB. Minimum configurations include 2× A100 80 GB, 2× H100 SXM5, one H200, one B200, or one B300; recommended configurations differ. These are Kolibri model requirements, not a universal platform minimum.
It applies to Kolibri’s downloadable model weights. Hosted inference, the platform, and enterprise applications are separate offerings with their own access terms.
Os dados de tráfego são apenas para referência.