A Hummingbird Lands on German Reunification Day

The timing was deliberate. On October 3, 2026, the Day of German Reunification, Aleph Alpha released Kolibri. Kolibri is German for hummingbird, and the name fits the engineering: a 78.1-billion-parameter model that only ever uses about 3.5 billion of them at a time.

The release landed on the company’s blog under a headline that made no attempt at subtlety: “Kolibri Has Landed: A Sovereign Open-Weight Model.” The full weights went up on Hugging Face the same morning, under the Apache 2.0 license. That last detail matters more than the poetry. Apache 2.0 means anyone can download Kolibri, run it, fine-tune it, and ship commercial products on top of it, without asking permission or filling out a form.

This is Aleph Alpha’s answer to the biggest question in European AI: can Europe build a serious model of its own, on its own terms?

The Engineering in Plain Terms

Kolibri is a mixture-of-experts model. Think of it as a model built from hundreds of specialists instead of one generalist. In each of its 50 layers, a router picks 6 experts out of 384 to handle the current token. Total parameters: 78.1 billion. Active on any given token: roughly 3.46 billion, or about 4.4 percent.

The point of that split is cost. All 78 billion parameters have to live in memory (about 78 GB in FP8, so the minimum hardware is two NVIDIA A100 80GB GPUs), but only a fraction of them burn compute on each token. You get the knowledge capacity of a giant model with the running cost of a much smaller one.

The context window is long: 262,144 tokens natively, extendable to a million through configuration. The model ships with four reasoning effort levels (none, low, medium, high) and tool calling. It handles English and German, with German making up about 21.3 percent of pre-training data, English around 62 percent, and code about 14 percent. Aleph Alpha even built a custom tokenizer, UniBPE, with a 128,000-token vocabulary tuned for German compound words, so German text costs fewer tokens to process.

Training itself is part of the pitch. Kolibri was trained on 768 NVIDIA B200 GPUs in Germany and Finland, under European and German law, on roughly 24 trillion tokens. Before committing to the full run, the team validated the pipeline on a smaller sibling, Kolibri Origin (30 billion total, 3 billion active, 65k context), then scaled the same approach up. Pre-training stayed stable across hardware failures and dropped connections without human intervention, which at this scale is not a small achievement. One faulty node at 768 GPUs usually means a dead run and a 3 a.m. pager.

The Benchmarks, Honestly Framed

Aleph Alpha publishes its numbers, as every lab does, and they should be read the way all vendor benchmarks are read: as a starting point, not a verdict.

Kolibri scores 96.9 on AIME 2025, 84.3 on GPQA Diamond, and 85.9 on LiveCodeBench v6, per the company’s reporting. The more interesting chart is not a ranking. It plots average score against decoded text per second per GPU, against Kolibri Origin, Qwen3.6-35B-A3B, Nemotron 3 Super, and Mistral Small 4. Aleph Alpha claims Kolibri sits on the Pareto frontier there: best quality for the serving cost, against models with up to four times its active parameter count.

Two details stand out. First, the company ran its math and science benchmarks in German and published that column alongside the English one. Almost nobody does this, and it is exactly what a model pitched at German public administration should be doing. Second, the model is signed to the EU General-Purpose AI code of practice, which is the compliance story European customers actually need to hear.

The model card also notes Aleph Alpha designed Kolibri to refrain from answering when it lacks supporting evidence, an anti-hallucination stance that matters for mission-critical use. Take it as a design goal to verify in practice, not a solved problem.

Sovereignty You Can Download

“Sovereign AI” gets thrown around a lot. Kolibri gives it a concrete meaning. Aleph Alpha uses the word in two senses: how the model was built (by teams in Germany, on infrastructure in Germany and Finland, under European and German law) and how it reaches customers (open weights they can run in their own data centers).

That second part is the real story. It is the same direction the enterprise market has been moving all month. IBM made its coding agent platform self-hostable this week, letting companies keep code and context inside their own walls. Kolibri takes the idea further down the stack: the model itself, downloadable, Apache-licensed, yours to run. No API key. No vendor with a kill switch. No sensitive documents traveling to someone else’s cloud.

“Kolibri demonstrates that we have the talent and expertise in Germany to develop competitive AI models,” said CEO Ilhan Scheer. “For us, AI sovereignty means freedom of choice by retaining the ability to build and advance this technology, and giving customers control over how they use it.”

Why It Matters Beyond Germany

The open-model conversation in 2026 has been dominated by the US and China. Reflection is reportedly about to ship the American answer. DeepSeek and Qwen set the bar the American models are chasing. Kolibri makes the map triangular: a European open-weight model, built under European law, benchmarked in German as well as English, and released under a license that lets businesses actually use it.

For regulated sectors, hospitals, banks, aerospace contractors, and government agencies across Europe, the calculation is simple. The best closed models are brilliant and unusable for your most sensitive data. Kolibri is the attempt to close that gap: frontier-adjacent quality, two GPUs of hardware, your building, your rules.

It will not be the biggest model of 2026. That is not the point. Kolibri is proof that a 200-person team in Heidelberg can ship a serious open-weight model on its own infrastructure, in months rather than years, and hand it to the world under Apache 2.0. The hummingbird landed. Watch what it builds next.