Aleph Alpha releases Kolibri-1 open-weight model under Apache 2.0

Overview

Updated 1 hour ago

On the Day of German Reunification, Aleph Alpha posted the full weights of its Kolibri-1 model to Hugging Face under an Apache 2.0 license. Anyone can now download, run, and modify the 78-billion-parameter English-German model without asking permission.

Kolibri is a mixture-of-experts model that activates only about 3.46 billion of its 78 billion parameters per token. Each request costs the compute of a small model while drawing on the capacity of a much larger one. It was built in Germany and trained on infrastructure in Germany and Finland, with the company controlling data, training, and evaluation end to end.

The release broadens the audience. Earlier rollout positioned Kolibri as a sovereign tool for German government and industry running their own infrastructure. Now any developer can use it, with one trade-off: the full 78-gigabyte model must stay in memory even though only a fraction activates per token.

Why it matters

EU agencies and companies now have an Apache-licensed, EU-trained model they can run on their own servers, without depending on US or Chinese AI.

Questions about this story

Free account needed to ask — your question is kept and asked for you right after sign-up. Answers are public.

No questions yet — be the first to ask.

Key Indicators

78.1B

Total model parameters

The full model, including all experts and shared layers.

3.46B

Parameters activated per token

About 4.4% of the model, selected by the routing mechanism for each token.

1,048,576

Maximum context length in tokens

Validated up to one million tokens; 262,144 is recommended for serving efficiency.

24T

Training tokens

Across pre-training, mid-training, and long-context extension phases.

392K

Pre-training GPU-hours

On 768 NVIDIA B200 accelerators over 21 days, excluding later training phases.

Voices

Curated perspectives — historical figures and your fellow readers.

Ever wondered what historical figures would say about today's headlines?

People Involved

Jonas Andrulis

Chief executive and co-founder of Aleph Alpha

Oversaw Kolibri's development and open-weights release

Organizations Involved

Aleph Alpha

AI research and product company

Model developer and publisher

German AI company that built Kolibri on EU infrastructure and released it under an open license.

Timeline

1 event Latest: Today

October 2026
Aleph Alpha releases Kolibri-1 on Hugging Face
Today Release

Full weights of the 78-billion-parameter English-German model published under Apache 2.0, letting anyone download, run, and modify it.
Today

Scenarios

German agencies run Kolibri in production

Possible Resolves by Oct 3, 2027

Discussed by: Aleph Alpha's own positioning and AI Act compliance coverage in trade press

German federal and state agencies handling sensitive workloads adopt Kolibri on their own infrastructure, drawn by EU AI Act compliance, Apache licensing, and a full-pipeline audit trail. The 78-gigabyte footprint fits two A100 or H100 cards, within reach of mid-size agencies.

Kolibri stays a niche sovereign tool

Possible Resolves by Oct 3, 2027

Discussed by: Observers noting the 78-gigabyte memory footprint and rapid US open-model releases

The full model must sit in memory on substantial hardware, limiting who can run it. US labs keep releasing larger open weights, and Kolibri stays confined to German public-sector contracts and research pilots without broad developer uptake.

Kolibri sparks a wave of EU sovereign open models

Possible Resolves by Oct 3, 2027

Discussed by: Coverage framing Kolibri as the first EU-native AI-Act-compliant open-weight model

Other European labs follow with their own sovereign open-weight models trained on EU infrastructure, turning sovereignty into a distinct category. The Apache license and compliance template set the pattern competitors must match.

Historical Context

3 moments from history that rhyme with this story — and how they unfolded.

September 2023

Mistral 7B release (September 2023)

French lab Mistral AI released a 7.3-billion-parameter model under the Apache 2.0 license, rivaling much larger models like Meta's Llama 2 on several benchmarks. It was the first major open-weights release from a European lab.

Then

Mistral 7B became a staple for on-premise and edge deployments and helped Mistral raise large funding rounds.

Now

It established European labs as credible open-weight players and made Apache-licensed models a normal route to market.

Why this matters now

Kolibri follows the same playbook, but adds sovereignty and EU AI Act compliance as the central selling point rather than raw capability.

December 2024

DeepSeek-V3 release (December 2024)

Chinese lab DeepSeek released V3, a 671-billion-parameter mixture-of-experts model that activates only 37 billion per token. It was trained for far less than most comparable Western models.

Then

DeepSeek matched frontier labs on benchmarks, and its app topped app-store charts in January 2025.

Now

The release reframed efficient architecture as a strategic lever, prompting Western investors to question high training costs.

Why this matters now

Kolibri uses the same MoE efficiency pattern, scaled for a different goal: low serving cost on EU infrastructure rather than headline scale.

July 2023

Meta's Llama 2 release (July 2023)

Meta released Llama 2, a family of models up to 70 billion parameters, under a permissive license allowing commercial use. Anyone could download and run it.

Then

A wave of fine-tuned variants and on-premise deployments followed, proving open weights could rival closed models.

Now

It reset the industry's default: major labs now treat open weights as a viable distribution channel.

Why this matters now

Kolibri's Apache release extends that open-access path, adding EU data-sovereignty guarantees that no US lab offers.