Kolibri-1: German business documents and an AI that should know when to abstain

When a supplier disputes a warranty claim, I do not want AI to invent a plausible warranty period. I want the document that establishes it. If the contract appendix is missing, the case should reach a person.

That makes Aleph Alpha’s Kolibri-1 worth examining. Released on 3 October 2026, it focuses on German and English, with weights available under Apache 2.0. The developer says it trained the model to abstain when the supplied context does not support an answer. That warrants a test; it is no guarantee of correctness. Aleph Alpha’s announcement.

Where I would start

My first pilot would serve a Czech company working with German suppliers. It would take emails, purchase orders and technical documents and produce a draft response linked to specific passages. Sending the email would require approval.

I would run the pilot in German. Czech and Slovak quality would need separate evaluation: success in one language is not evidence of success in another. Translating inputs also introduces another opportunity to distort a technical term.

This builds on control over enterprise AI deployment. The distinct question here is how a newly released model handles missing evidence in a concrete document workflow.

Active parameters do not size your server

The model card lists roughly 78 billion total parameters and 3.46 billion active per token. FP8 weights occupy about 78 GB. Maximum context is 1,048,576 tokens, but the developer recommends no more than 262,144 for serving efficiency and complex tasks. Deployment uses Aleph Alpha’s vLLM plugin.

My budgeting rule follows from that distinction: do not size infrastructure from the active parameter count. Include memory for concurrent requests, document length, spare capacity and the operator’s time. Choose the GPU configuration after measuring the intended workload.

A proposed n8n workflow

This is a pilot design, not an integration tested for this article:

1. Internally hosted n8n receives an email and assigns a case ID. OCR for scanned attachments remains a separate step.
2. Retrieval returns passages the user is permitted to access, each with a document ID and version.
3. The model proposes an answer, supporting sources and missing information. The application validates the required output structure.
4. A separate check confirms that cited passages exist. A reviewer checks whether they actually support the conclusion; a valid reference alone is insufficient.
5. The draft enters an approval queue. Running the same case ID again must not create another case or resend an email.

If keeping data inside the company is the objective, review OCR, retrieval, logs and backups too. A local model does not control the document’s path through other services.

Measure excessive abstention too

I would prepare 60 manually labelled cases: 20 with complete evidence, 20 missing a decisive fact and 20 containing conflicting documents. These numbers describe a small starting set, not a statistical guarantee.

Track unsupported answers, unnecessary refusals on answerable cases, citation correctness, correction time and cost per approved result. A model that refuses every request can look cautious while being useless to the team.

Run it alongside the existing solution, without automatic CRM or ERP writes. Expand only if it reduces corrections at acceptable latency and operating cost. For this business, Kolibri’s value will be demonstrated by an answer grounded in a German source document, not by context capacity alone.

Sources: original announcement, 3 October 2026, model card and deployment instructions. I did not run an independent model benchmark for this article.