Kolibri 1
Aleph Alpha's English-German reasoning model: 78B mixture of experts with only 3.5B active, 1M context, Apache 2.0.
In short. Kolibri 1 is a German-built AI model that is fluent in both German and English. It is designed for organizations that want to run AI on their own servers, such as public agencies, factories and aerospace firms. It is large on disk but fast to run, and the Apache 2.0 license allows commercial use.
Summary
Aleph Alpha released Kolibri 1 on 3 October 2026 under the Apache 2.0 license. It is a bilingual German and English mixture-of-experts reasoning model with 78 billion parameters in total and about 3.5 billion active per token, with explicit reasoning and tool calling. The company says it was built in Germany, trained on infrastructure in Germany and Finland, and designed with the EU AI Act, the GPAI Code of Practice and the GDPR in mind.
The model has a 1,048,576-token context window, although the card recommends keeping prompts to 262,144 tokens or fewer for serving efficiency, since 262,144 is the length it was natively trained on. It has 50 layers with a 4:1 mix of sliding-window and global attention, and 384 experts per layer, of which 6 routed experts plus 1 shared expert are used per token. It was pre-trained on 20 trillion tokens.
Weights are published in FP8, about 79 GB, and in BF16, about 156 GB. The card lists a minimum of one H200, B200 or B300 GPU, or two 80 GB GPUs such as A100 or H100, for the FP8 version.
Aleph Alpha says it was trained to answer "I don't know" when the supplied documents do not contain the answer, which matters for document-grounded work. It reports 96.9 on AIME 2025 and 84.3 on GPQA Diamond, and says it matches models with up to four times as many active parameters, such as Nemotron 3 Super. These are the company's own results.
Good for
- German-language assistants and document question answering, including retrieval-augmented generation
- Public-sector, industrial and other regulated deployments that want a European-built model on their own servers
- Agents with tool calling that must be cheap to run, since only about 3.5B parameters are active per token
- Long-document work with a permissive Apache 2.0 license
Watch out for
- Only German and English are supported; it is not a broad multilingual model
- Text only, with no image, audio or video input
- Even in FP8 it needs a data-center GPU such as an H200, or two 80 GB cards
- The benchmark comparisons are Aleph Alpha's own; test on your own German documents
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Estimates for the Kolibri 1 FP8 variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.
| Variants | Parameters | BF16 | FP8 / INT8 | 4-bit (Q4) | Fits on |
|---|---|---|---|---|---|
| Kolibri 1 FP8 FP8 | 78B 3.5B active | 179 GB | 90 GB | 54 GB | Workstation |
| Kolibri 1 BF16 BF16 | 78B 3.5B active | 179 GB | 90 GB | 54 GB | Workstation |
Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.
The license in plain words
One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.
See also
Frequently asked questions
- What is Kolibri 1?
- Kolibri 1 is a German-built AI model that is fluent in both German and English. It is designed for organizations that want to run AI on their own servers, such as public agencies, factories and aerospace firms. It is large on disk but fast to run, and the Apache 2.0 license allows commercial use.
- Can I use Kolibri 1 commercially?
- Yes. Kolibri 1 is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is Kolibri 1 open source?
- Kolibri 1 is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run Kolibri 1 on my own computer?
- Only on a workstation. The smallest variant, Kolibri 1 FP8, needs about 54 GB of memory at 4-bit precision, so plan for a 48 to 96 GB GPU, several graphics cards, or a Mac with plenty of memory.
- When was Kolibri 1 released?
- Kolibri 1 was released by Aleph Alpha on October 3, 2026.
- How long a context window does Kolibri 1 have?
- Kolibri 1 supports a context window of up to 1M tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- Aleph Alpha: Kolibri Has Landed, a Sovereign Open-Weight Model (3 Oct 2026) aleph-alpha.com
- Kolibri-1 model card (Hugging Face) huggingface.co
Release date and facts last verified November 8, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| gpt-oss OpenAI | 21B – 117B 5.1B active | 128K | Apache 2.0 (Permissive) | 15 GB Laptop or phone | |
| Nemotron 3.5 NVIDIA | 8B – 120B 12B active | 1M | OpenMDW-1.1 (Permissive) | 5.52 GB Laptop or phone | |
| Granite 4.2 IBM | 3.7B – 29.3B | 128K | Apache 2.0 (Permissive) | 2.55 GB Laptop or phone | |
| Nex-N2.6 Nex-AGI | 35.1B – 1.6T | 256K | Apache 2.0 (Permissive) | 24 GB Gaming GPU |