TL;DR
Get tech for your team delivered free — and shop member deals
- Fast, free delivery on millions of items
- Access to Prime Big Deal Days deals on October 6–7
- Prime Video, Amazon Music and more included
Aleph Alpha released Kolibri on Oct. 3, 2026, as an open-weight German-English model under the Apache 2.0 license. Its technical report describes a 78.1-billion-parameter mixture-of-experts model that activates about 3.46 billion parameters per token; performance claims come from the company’s evaluations, and independent results were not provided in the source material.
Kolibri was trained from scratch using about 24 trillion tokens and 768 NVIDIA B200 graphics processing units, according to Aleph Alpha’s technical report and model card. More than a fifth of its training tokens were German, the report says. The model’s stated knowledge cutoff is June 18, 2026.
The model has a stated native context length of 262,144 tokens; the model card says it has been tested up to 1,048,576 tokens. It supports four reasoning settings—none, low, medium and high—and tool calling. At 8-bit floating point, its weights require about 78 GB of memory, according to the source material.
Aleph Alpha says its evaluations place Kolibri above every model it compared with in its size category in both German and English. That is a company-reported result: the source material does not include independent benchmark results or enough detail here to verify the comparisons. The weights are hosted on Hugging Face, while Aleph Alpha retains rights to its training code and methods.
A European Option for Language Models
Kolibri adds an openly licensed model built for German as well as English to a field where many widely used systems have been developed elsewhere. Organizations that need to process sensitive material may value the option to run its weights on their own infrastructure, rather than send data to an external hosted service. Whether that offers a practical advantage depends on their hardware, security arrangements and deployment needs.
Aleph Alpha describes the model as sovereign, pointing to its development by teams in Germany and training on infrastructure in Germany and Finland under German and European law. That positioning may be relevant to public agencies and businesses with data-governance requirements. It is the company’s characterization, however, and does not by itself establish that every use of the model will meet a particular organization’s legal or compliance obligations.
The mixture-of-experts design also makes the model’s efficiency claims more specific than a simple parameter count. Only a fraction of its parameters are active for each token, which can reduce computation compared with a dense model of similar total size. But operators still need to hold the full set of weights in memory, so the model is not equivalent to a small model in hardware requirements.
As an affiliate, we earn on qualifying purchases.
How Kolibri Was Built
Kolibri’s architecture is a mixture of experts: a router directs each token through selected specialist sub-networks instead of sending every token through all model parameters. Aleph Alpha reports 50 layers, each with 384 routed experts and one shared expert. Six routed experts are selected for each token, alongside the shared expert, yielding roughly 3.46 billion active parameters out of 78.1 billion in total.
The company also developed a tokenizer with a vocabulary of 128,000 tokens, using an approach it calls UniBPE. The technical report says it requires 11.2% fewer tokens for German text than GPT-5’s tokenizer among nine alternatives evaluated. This is a reported comparison by Aleph Alpha, not an independently verified result in the material provided. The design addresses a feature of German: compound words can be split into many pieces by tokenizers trained mainly around English text.
The model’s European development does not mean that all tools or data sources were European. Its model card says Google’s Gemma 4 was used to rephrase English web text, Mistral-NeMo to rephrase German text, and Qwen3-32B to label data used for quality filters. Aleph Alpha says it filtered training data for political bias and has signed the European Union’s General-Purpose AI Code of Practice.
“Teams built the model in Germany, trained it on infrastructure in Germany and Finland, under European and German law, with no foreign control.”
— Aleph Alpha, in its launch materials
Independent Testing Still Needed
The available source material does not provide independent evaluations of Kolibri’s German or English quality, nor enough benchmark detail to assess how Aleph Alpha selected comparison models and tasks. The company’s claim that Kolibri scores above every compared model of its size should therefore be treated as a company-reported evaluation, not a settled ranking.
It is also unclear from the provided material how the model performs across real-world tasks, how its long-context capability compares with other systems, or what hardware and operating costs users will face in practice. The 78 GB figure is an estimate for 8-bit weights and does not describe all memory needed for a deployment. The source does not give pricing for support or hosted services.
Apache 2.0 covers the model weights and configuration files, according to the report; Aleph Alpha retains rights to training code and methods. The precise terms and practical limits of using Kolibri should be checked against the published license and accompanying files. Open weights also do not independently settle questions about training-data provenance, model behavior or legal compliance in a specific application.
Deployment And Evaluation Ahead
The immediate next step is for developers and organizations to download the weights from Hugging Face, review the model card and license, and test the system against their own language, accuracy and security requirements. Those tests can help determine whether its German tokenizer, tool calling and context window suit a particular workload, and whether available hardware can support the full model.
Further independent benchmarks and documented deployment results would help clarify how Kolibri compares with other German-English models on common tasks, as well as the trade-offs between active computation and memory use. Aleph Alpha’s release materials establish the model’s availability and stated specifications; real-world performance and adoption remain to be seen.
Key Questions
What is Kolibri?
Kolibri is Aleph Alpha’s open-weight large language model for German and English, released on Oct. 3, 2026.
What license does Kolibri use?
Aleph Alpha says the model weights and configuration files are available under the Apache 2.0 license. The company retains rights to its training code and methods.
How large is the model?
Kolibri has 78.1 billion total parameters, with about 3.46 billion active for each token according to Aleph Alpha’s technical materials. At 8-bit floating point, the weights require about 78 GB of memory.
Can Kolibri run on an organization’s own servers?
Aleph Alpha describes Kolibri as deployable on customer-controlled infrastructure. The full weights still need to fit in memory, and the source material does not specify a single minimum hardware configuration for all deployments.
Are Kolibri’s benchmark claims independently verified?
The claim that it outperforms every compared model of its size in German and English comes from Aleph Alpha’s own evaluation. Independent results were not included in the source material.
Source: hn
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
