# Aleph Alpha Releases 78B Kolibri Open-Weight Model

Published 2026-10-03 · AI.info Team · AI.info

> Aleph Alpha released Kolibri, a German-English open-weight model with 78 billion total parameters and an Apache 2.0 license. Kolibri supports context windows up to one million tokens, though its model card recommends shorter contexts for many deployments.

> “Kolibri is an English-German Mixture-of-Experts Transformer with 78B total parameters, 3B active.”
>
> [Aleph Alpha](https://aleph-alpha.com/en/blog/kolibri-has-landed-a-sovereign-open-weight-model/)

Aleph Alpha released Kolibri on October 3, 2026, making the model’s full weights available on Hugging Face under the Apache 2.0 license. The German company is targeting organizations that want to run language models on their own infrastructure, particularly in public administration, industry and aerospace. Kolibri supports English and German, reasoning modes, and tool calling.

## Kolibri puts German and English first

The model is built for two languages rather than broad multilingual coverage. German makes up 21.3% of Kolibri’s pre-training tokens and Aleph Alpha developed a bilingual tokenizer to handle German word structure. The company also used translated data sparingly, aiming to give the model exposure to organic German-language material.

Kolibri follows an earlier internal model, Kolibri Origin. The new model grew from 30.6 billion total parameters and a 65,536-token trained context in Origin to 78.1 billion parameters and a 262,144-token trained context. It was trained on 20 trillion pre-training tokens, selected from more than 200 trillion raw tokens that the company filtered and curated.

## Low active compute, substantial memory needs

Kolibri is a mixture-of-experts model: it activates 3.46 billion parameters per token while retaining 78.1 billion in total. That design can limit computation for each generated token, but the full model still needs to fit in memory. Its [Hugging Face model card](https://huggingface.co/Aleph-Alpha/Kolibri-1-BF16) lists a roughly 156GB memory footprint for BF16 weights and a minimum configuration of four A100 80GB GPUs, four H100 SXM5 GPUs, two H200 GPUs, or one B200 or B300.

Aleph Alpha advertises context lengths up to 1,048,576 tokens, but that ceiling comes with a qualification: the model card recommends contexts of no more than 262,144 tokens for serving efficiency and complex tasks. The model was trained to 262,144 tokens, with longer contexts validated by the company up to one million. The difference matters to operators weighing a long-context feature against hardware and latency costs.

## Company benchmarks show strengths and gaps

Aleph Alpha’s published results show Kolibri scoring 96.0 on AIME 2026 in English and 90.0 on the German version. In the company’s comparison, those scores top the other listed models on both rows. The same table reports 21.5 on English Humanity’s Last Exam, below the 35.6 score listed for Qwen3.8, a model with 27 billion active parameters.

The results come from Aleph Alpha’s own evaluations, not an independent assessment. The benchmark comparisons use Aleph Alpha’s evaluation harnesses and, where applicable, the highest reasoning effort for each model. That gives customers a set of results to inspect, but the published scores should be read as vendor-reported comparisons rather than a settled measure of performance across deployments.

## A sovereignty pitch with a practical test

Aleph Alpha built Kolibri in Germany and trained it on infrastructure in Germany and Finland, under European and German law. It presents downloadable weights and on-premises deployment as ways for customers to keep data within their own systems and retain control over model use. Those are company claims about how the model was built and deployed, not a guarantee that every customer’s setup meets its own legal or security requirements.

Kolibri’s clearest trade-off is visible in its specifications: only a small fraction of its parameters activate for each token, but all of its BF16 weights still require substantial memory. The model offers organizations a German-English option they can download and run themselves; the listed minimum for its BF16 version is four A100 80GB GPUs.

## Sources

1. Original source · Blog post [Aleph Alpha](https://aleph-alpha.com/en/blog/kolibri-has-landed-a-sovereign-open-weight-model/) aleph-alpha.com

Other pages read

2. [Hugging Face](https://huggingface.co/Aleph-Alpha/Kolibri-1-BF16) huggingface.co

## Related stories

- [Cohere and Aleph Alpha Sign Definitive Merger Agreement](https://ai.info/news/cohere-aleph-alpha-sign-definitive-merger-agreement)
- [Aleph Alpha Finds Party-Line Responses in Chinese LLMs](https://ai.info/news/aleph-alpha-chinese-llms-party-line-responses)
- [Mistral raises €3B to build Europe’s sovereign AI stack](https://ai.info/news/mistral-raises-3b-to-build-europes-sovereign-ai-stack)
- [Fujitsu Sets November Sales for MONAKA CPU and Sovereign AI Server](https://ai.info/news/fujitsu-monaka-cpu-sovereign-ai-server-november-sales)
- [Cohere targets C$7 billion in sovereign AI funding round](https://ai.info/news/cohere-targets-cad-7-billion-sovereign-ai-funding)

## Explore

[More articles](https://ai.info/news)

---

Canonical: https://ai.info/news/aleph-alpha-kolibri-78b-open-weight-model
