# Kolibri gives teams another open-weight AI option. Here’s what it takes to run it.

Agent contract: 1.2.0

Germany’s Aleph Alpha released a downloadable German-English model on October 3. It offers another route to self-hosted AI, while showing why open weights, open-source development and easy local use are different things.

Canonical: https://brightaifuture.com/discoveries/kolibri-open-weight-control
Format: discovery
Source publication: 2026-10-03
Bright publication: 2026-10-04
Substantive update: None recorded
Evidence and review: Emerging; confidence: unassessed; approved; ai-assisted. AI-assisted source comparison of the October 3 announcement and current model card, separate inference license, technical report and data summary. Original Bright analysis concerns deployment choice and practical constraints. Vendor benchmarks, million-token validation and sovereignty claims are not independently reproduced performance or legal guarantees.

## Another choice for teams

Aleph Alpha has [released Kolibri 1](https://aleph-alpha.com/en/blog/kolibri-has-landed-a-sovereign-open-weight-model/), a model designed for German- and English-language tasks, including document analysis, question answering and tool use. Its [weights are available on Hugging Face](https://huggingface.co/Aleph-Alpha/Kolibri-1) under Apache 2.0 terms. A team with the right infrastructure can download the model and run it on systems it controls rather than relying only on a hosted model service.

That choice matters for organizations working with internal documents or building German-language tools. They can test Kolibri on their own questions, decide where the model runs and adapt the released weights to a specific workflow. Those are practical options created by the release itself. Whether Kolibri is the best choice for a particular workload will take testing beyond the launch materials.

## Small active count, substantial hardware

Kolibri uses a mixture-of-experts design. It has about 78 billion parameters in total, but uses roughly 3.46 billion for each token it processes. Think of it as calling on a small group of specialists for each step rather than asking every specialist to work every time. That can reduce computation per token. The other specialists do not disappear from memory: the full model still has to be stored while it runs.

That is the first practical limit to understand. The [model card](https://huggingface.co/Aleph-Alpha/Kolibri-1) lists a memory footprint of about 78 GB for its FP8 weights. Its official minimum hardware examples include two 80 GB A100 GPUs, two H100 SXM5 GPUs, or one H200, B200 or B300. Memory for the weights is only part of a deployment; serving longer documents or multiple users adds further demands. This is an option for teams able to provision serious GPU capacity, not a promise that the full release will run well on an ordinary laptop.

## A context ceiling with a qualification

The company also says Kolibri can support up to 1,048,576 tokens of context. That is a potentially useful ceiling for long documents, but it needs a qualifier: Kolibri’s final long-context training went to 262,144 tokens, and Aleph Alpha recommends staying at or below that length for serving efficiency and complex tasks. The million-token figure is the company’s validation claim, not a guarantee that every question over a very long input will be answered reliably.

## What is open, and what is not

The word “open” has a boundary here too. Kolibri’s model card says its Apache 2.0 license covers the weights and configuration files in the repository. Aleph Alpha has also released a separate [Apache-licensed inference plugin](https://github.com/Aleph-Alpha/aleph-alpha-inference) and a detailed [technical report](https://aleph-alpha.com/downloads/tech-report.pdf). But the model card explicitly excludes other artifacts, including underlying code and training methods, from the model-repository license. It provides a [summary of training-data sources](https://aleph-alpha.com/downloads/data-summary.pdf), not the complete training corpus. “Open-weight” is the precise description; a downloadable model is not automatically a fully reproducible open-source AI system. Bright’s [open-model guide](https://brightaifuture.com/open-models#openness) explains that distinction in more detail.

## The questions that make the release useful

Aleph Alpha reports strong results in its own benchmark comparisons and positions Kolibri as a “sovereign” model for organizations that want more control over deployment. Those are the company’s claims, not independently reproduced performance or a blanket legal-compliance guarantee. The model card itself calls for people to review outputs before consequential actions.

The useful news is concrete: there is now another English-German checkpoint that qualified teams can download, inspect and test on their own workloads. The next questions are equally concrete. Does it answer their documents accurately, at an acceptable speed and cost, on hardware they can actually use? The release makes those questions testable. It does not answer them for every team.

## The human problem

Teams working with internal documents or building German-language tools need to choose where AI runs, evaluate answers on their own material and understand the practical cost of control.

## The prior constraint

A hosted model service is one route to those tools. Downloadable weights create another, but teams still need compatible inference software, enough GPU memory and evidence that the model works for their actual questions.

## AI’s actual role

Kolibri is a German-English mixture-of-experts language model for document analysis, question answering, reasoning and tool use. About 3.46 billion parameters are active per token out of 78.1 billion total; the full model must remain in memory.

## The documented result

Aleph Alpha released Kolibri 1 on October 3, 2026 with downloadable weights and configuration files under Apache 2.0. The model card lists about 78 GB of FP8 weight memory and server-class minimum GPU configurations. It recommends contexts of at most 262,144 tokens for efficient serving and complex tasks, while claiming validation up to 1,048,576 tokens. The inference plugin is separately Apache-licensed.

## Why it may matter

Bright’s analysis: the release gives qualified teams another checkpoint they can download, inspect, adapt and test on systems they control. Choice of deployment can be useful for internal-document and German-language work. Its value depends on accuracy, latency and total serving cost on each team’s workload; the release does not establish those outcomes for every organization.

## Limitations

Open-weight: the model-repository Apache 2.0 grant covers weights and configuration files, not a complete training stack or corpus. The separate inference plugin has its own license.

The roughly 3.46B active parameter count does not mean a 3.46B memory footprint. The model card lists about 78 GB of FP8 weights before serving overhead and server-class GPU examples.

The million-token figure is company-validated extrapolation beyond the 262,144-token final training length; Aleph Alpha recommends at most 262,144 for complex tasks and serving efficiency.

Benchmark comparisons and sovereignty benefits are vendor claims. Bright has not reproduced performance or established a blanket legal-compliance guarantee; outputs need human review before consequential action.

## Unresolved questions

How accurately does it answer a team’s own documents, including questions with no supported answer?

What speed and total serving cost does it achieve on hardware the team can actually provision?

Which independent evaluations reproduce the reported performance and long-context reliability?

## Provenance and history

{
  "dates": {
    "eventDate": "2026-10-03",
    "publicationDate": "2026-10-03",
    "captureDate": "2026-10-04",
    "lastReviewedDate": "2026-10-04"
  },
  "provenance": {
    "origin": "editorial",
    "externalId": "https://aleph-alpha.com/en/blog/kolibri-has-landed-a-sovereign-open-weight-model/"
  },
  "revisions": [
    {
      "id": "revision:5b063b167b9ae28a62ad",
      "recordedAt": "2026-10-04",
      "summary": "Kolibri gives teams another open-weight AI option. Here’s what it takes to run it.",
      "sourceIds": [
        "source:kolibri-release-october3",
        "source:kolibri-model-card",
        "source:kolibri-inference-plugin",
        "source:kolibri-technical-report",
        "source:kolibri-data-summary"
      ]
    }
  ],
  "corrections": []
}

## Original sources

- [Aleph Alpha · Kolibri Has Landed · October 3, 2026 · company release announcement](https://aleph-alpha.com/en/blog/kolibri-has-landed-a-sovereign-open-weight-model/)
- [Aleph Alpha · Kolibri 1 model card · weights, hardware, context, intended use and license scope](https://huggingface.co/Aleph-Alpha/Kolibri-1)
- [Aleph Alpha inference plugin · separate Apache 2.0 inference software](https://github.com/Aleph-Alpha/aleph-alpha-inference)
- [Aleph Alpha · Kolibri technical report · company-reported training and evaluations](https://aleph-alpha.com/downloads/tech-report.pdf)
- [Aleph Alpha · public summary of Kolibri training-data sources](https://aleph-alpha.com/downloads/data-summary.pdf)

## Continue exploring

- [Your AI conversation feels private. Where it runs matters.](https://brightaifuture.com/discoveries/ai-conversation-privacy)
- [Open Models](https://brightaifuture.com/open-models)
- [Open intelligence](https://brightaifuture.com/worlds/open)
- [What changes when powerful models become open-weight?](https://brightaifuture.com/threads/open)