Aleph Alpha Kolibri: Germany's New Open-Source AI Model Puts Europe Back in the AI Race

Aleph Alpha Kolibri suddenly launched and nobody expected it

Many Europeans were disappointed when Mistral recently shifted its focus toward providing users with access to Chinese open-source models, suggesting a slower pace for its own releases. To some, this felt like the end of Europe's frontier ambitions. The gap is steadily growing and is already quite big as Chinese and U.S. labs release new frontier models every one and a half to two months, while Mistral's latest models are about six months old, and the platform prioritizes newer, better, open-weight models from China.

Yet Europe is not out of the race. Unexpectedly, the German startup Aleph Alpha just released Kolibri, a powerful, compact model under the open Apache 2.0 license with full weights on Hugging Face.

GERMANY JUST RE-ENTERED THE AI MODEL RACE.

Aleph Alpha builds AI mainly for enterprise and public-sector customers, so it does not offer a consumer chat platform itself. You can still access Kolibri and chat for free via xPrivo from day one, thanks to the partnership with Tesseracted.

What is Aleph Alpha Kolibri?

Kolibri (German for "hummingbird") is a Mixture-of-Experts Transformer with 78 billion total parameters, of which only about 3 billion are active per token. This makes it fast and inexpensive to run while retaining the capacity of a much larger model. The Aleph Alpha team also chose a special day to release the model: October 3, 2026, the Day of German Unity.

• Architecture: custom MoE with 384 experts per layer (6 routed plus 1 shared), 50 layers, and sliding-window attention with full attention in every fifth layer.
• Context window: up to 1 million tokens (262k natively, which is the recommended range for efficiency).
• Training: about 24 trillion tokens in total on 768 NVIDIA B200 GPUs, on infrastructure in Germany and Finland.
• Languages: built for German and English. 21.3% of pre-training tokens are German, and a custom tokenizer handles German compound words efficiently, for example "Bundessozialgerichtes" becomes Bundes | sozial | gericht | es.
• Reasoning modes: four effort levels (none, low, medium, high) to trade cost and speed against answer quality.
• Fewer hallucinations: trained to say "I don't know" when the answer is not in the provided documents, using Aleph Alpha's Merlin-Arthur method.
• License: Apache 2.0, with weights on Hugging Face.

Aleph Alpha designed Kolibri with the EU AI Act, the GPAI Code of Practice, and the GDPR in mind. This emphasis on sovereignty and control is precisely what European companies need, which is why it is also a good fit for a privacy-first European assistant like xPrivo.

How to try Aleph Alpha Kolibri for free

In case you didn't know, xPrivo is an open-source, privacy-first AI assistant from Europe that runs exclusively on European infrastructure by default. Using Kolibri is easy. It takes only three steps:

1. Open xPrivo Chat.
2. Open the model list and select Kolibri.
3. Start chatting in German or English. It is free to use right now.

How to use the Aleph Alpha Kolibri API

Our day-one launch is possible thanks to our partnership with Tesseracted, a German company that brings Kolibri to privacy-conscious users and teams.

Want to build Kolibri into your own project? Tesseracted offers an early Aleph Alpha Kolibri API, so you can integrate the model into your app, workflow or internal tool without running the infrastructure yourself. Request early API access at Tesseracted.

Prefer to self-host? The weights are on Hugging Face as Aleph-Alpha/Kolibri-1. The FP8 model needs roughly 78 GB of memory (for example 2x H100 or 1x H200 or B200) and is served with vLLM.

How does Kolibri compare? The numbers

Aleph Alpha compares Kolibri with three open-weight models of similar or larger size: Qwen3.6-35B-A3B (Alibaba), Nemotron 3 Super 120B-A12B (NVIDIA) and Mistral Small 4 119B-A6B (Mistral). With only about 3B active parameters, Kolibri matches or beats models with up to four times as many.

Benchmark	Kolibri	Qwen3.6 35B-A3B	Nemotron 3 Super	Mistral Small 4
Overall (EN)	75.5	71.4	73.0	63.1
Overall (DE)	70.8	67.3	67.9	61.4
AIME 2025 (math)	96.9	84.6	91.7	79.8
AIME 2026 (math)	96.0	91.0	90.4	83.1
GPQA Diamond (knowledge)	84.3	83.4	78.0	74.7
LiveCodeBench v6 (code)	85.9	82.5	82.0	71.2
Humanity's Last Exam	21.5	21.1	20.6	9.7
Tau2-Bench Telecom (agentic)	94.7	99.1	68.1	41.5
BFCL v4 (tool calling)	61.4	67.2	61.0	58.0
AA-LCR (long context)	68.3	69.7	67.0	52.3

Source: Aleph Alpha's own benchmark runs (Hugging Face model card, highest reasoning effort). Vendor benchmarks are a starting point, so test on your own tasks.

Where Kolibri wins: math, science knowledge, coding and German-language performance. It also leads the compared models on Aleph Alpha's agentic-RAG benchmark Honeypot, and it avoids wrong answers far more often than its predecessor: 44% of AA-Omniscience items, up from 15% for Kolibri Origin.

Where it doesn't: Qwen3.6 leads in several agentic, tool-calling and long-context tests. And these three comparison models are from the spring. Newer open-weight models are stronger, as the next table shows.

Kolibri in the open-weight landscape

* Estimate. Kolibri is not yet listed on Artificial Analysis, so the value is an approximation based on its performance relative to Qwen3.6. All other scores: Artificial Analysis Intelligence Index.

To be honest, Kolibri is neither a trillion-parameter frontier model nor the top global open-weight model. However, it offers a rare combination: a fully open, European-built model that is strong in German and efficient enough to run on two GPUs. It is clearly ahead of older European and US open models in its size class which is already a big win.

Europe is not done yet

Kolibri demonstrates Europe's ability to develop competitive models from scratch. Aleph Alpha developed the data pipeline and the pre- and post-training processes in-house. They advanced from the unreleased 30B Kolibri Origin to the 78B Kolibri model in just three months. This development comes at an interesting time, as Aleph Alpha has signed an agreement to merge with the Canadian AI company Cohere. Heidelberg will remain a research hub.

Kolibri is more than just a successful model release; it is also a powerful demonstration of Europe's commitment to open-source AI. While not a cutting-edge model, it demonstrates Europe's ability to develop competitive AI and share it openly. More may be on the way, as Mistral recently raised 3 billion euros in new funding led by Samsung. CEO Arthur Mensch says its next-generation model will significantly close the gap with U.S. AI labs like OpenAI and Anthropic. That is a big promise, and the benchmarks will have to prove it. Nonetheless, Europe is finally back in the AI race, and we can't wait to see what's next.

Aleph Alpha Kolibri FAQ

What is Aleph Alpha Kolibri?

A German open-weight Mixture-of-Experts language model with 78B total and about 3B active parameters, released on 3 October 2026 under Apache 2.0.

Is there an Aleph Alpha Kolibri Chat?

Yes. You can chat with Kolibri for free on xPrivo by selecting it in the model list.

Is there an Aleph Alpha Kolibri API?

Our German partner Tesseracted offers an early Kolibri API. You can also self-host the weights from Hugging Face with vLLM.

Is Kolibri really open source?

The weights are released under Apache 2.0. Aleph Alpha keeps its rights to training code, architecture and methods. The best term for this is "open-weight".

Sources

• Aleph Alpha: Kolibri Has Landed: A Sovereign Open-Weight Model https://aleph-alpha.com/en/blog/kolibri-has-landed-a-sovereign-open-weight-model/
• Hugging Face: Aleph-Alpha/Kolibri-1 model card: https://huggingface.co/Aleph-Alpha/Kolibri-1