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, an open-weight German- and English-language model, on Oct. 3, 2026, under the Apache 2.0 license. The company reports 78.1 billion total parameters, with about 3.46 billion active for each token, and says its evaluations place Kolibri ahead of compared models of similar size in both languages. Independent results and practical deployment requirements remain to be established.
Aleph Alpha released Kolibri, an open-weight large language model for German and English, on Oct. 3, 2026, making its weights and configuration files available under the Apache 2.0 license. The company describes it as a model built with European deployment and regulatory requirements in mind; its reported benchmark advantage over similar-sized models is based on Aleph Alpha’s own evaluation.
Kolibri is a mixture-of-experts model with 78.1 billion parameters in total. Aleph Alpha says about 3.46 billion parameters are active per token, or roughly 4.4% of the full parameter count. That design can limit the computation used for each token compared with activating the full network, but it does not eliminate the need to store the full model: the model card says the complete set of weights must be held in memory.
Aleph Alpha’s technical report lists a 262,144-token native context window, with tests extending to 1,048,576 tokens. It says the model was trained on about 24 trillion tokens, more than one-fifth of them in German, using 768 NVIDIA B200 GPUs. The stated knowledge cutoff is June 18, 2026. The weights require about 78 GB at 8-bit floating point, according to the report, so deployment still involves substantial hardware and memory requirements.
The company says Kolibri was trained from scratch on infrastructure in Germany and Finland, and that its design took the EU AI Act into account from the outset. The model supports tool calling and four reasoning settings: none, low, medium and high. The model is available on Hugging Face; Aleph Alpha retains rights to its training code and methods, even as the weights and configuration files use Apache 2.0.
European Control and Deployment Choices
Kolibri adds an openly licensed option for organisations seeking a model they can inspect, adapt and run on their own infrastructure. That may appeal to public agencies, companies handling sensitive records, or businesses that need more control over where data is processed. The license and availability of weights give deployers more freedom than a service accessible only through a vendor’s hosted interface, though the model’s size means local operation is not automatically simple or inexpensive.
Aleph Alpha calls this approach “sovereign”, describing both the model’s European development and the customer’s ability to control deployment. The term does not mean every component or influence is European: the model card says some training material was generated or labeled using Google’s Gemma 4, Mistral-NeMo and Qwen3-32B. Nor does a European development location alone establish compliance for every use. Organisations will still need to evaluate the model, its data handling and their own deployment against applicable legal and operational requirements.
The release also matters for German-language performance. German compounds can be split inefficiently by tokenizers developed around English text, increasing the number of tokens needed to process the same material. Aleph Alpha reports that Kolibri’s tokenizer uses 11.2% fewer tokens for German text than GPT-5’s tokenizer among the nine alternatives it tested. That is a company-reported comparison, not evidence by itself that Kolibri is more accurate or less costly in every task.
high-performance GPU for AI training
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
How Kolibri Handles German Text
Kolibri’s architecture combines a large collection of specialist sub-networks with a router that selects a subset for each token. The technical report describes 50 layers, each containing 384 routed experts and one shared expert; six routed experts are selected for a token. This mixture-of-experts arrangement explains the difference between the model’s total parameter count and the smaller number used in each token’s computation. It also means that lower per-token computation should not be confused with a small memory footprint.
The tokenizer is another part of the German-language design. It has a vocabulary of 128,000 tokens and uses an approach Aleph Alpha calls UniBPE, combining byte-pair merging with a different rule for choosing merges. In the company’s example, the German word “Bundesverfassungsgericht” is split into six tokens by the cited GPT tokenizer and two by Kolibri’s. The example illustrates tokenization, not a direct measure of model quality.
The launch fits Aleph Alpha’s stated focus on models developed under European law and intended for controlled deployment. The company says it signed the EU General-Purpose AI Code of Practice. Its documentation also identifies non-European tools used in preparing some training data, a relevant qualification to broad claims of sovereignty. Kolibri’s release materials include a technical report, model card and launch post; the available source account draws on those documents as well as the author’s own tokenizer experiment.
“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 post
large memory server for machine learning
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Independent Results and Costs
The supplied release information does not establish how Kolibri performs in independent evaluations, or whether Aleph Alpha’s reported lead over models of similar size holds across different benchmarks, prompts and real-world tasks. The comparison results are the company’s own, and the available material does not provide enough detail here to assess every benchmark’s design or testing conditions.
It is also unclear how well the model performs across German and English beyond the reported evaluations, including on specialist material, factual accuracy, safety and multilingual use outside those two languages. The long context figures are reported capabilities; the available information does not show how consistently quality is maintained near the maximum context length.
Deployers will need to test hardware costs, response speed and reliability in their own settings. The model’s 78.1-billion-parameter footprint remains relevant even though fewer parameters are active per token. The source material also does not settle how training-data provenance, the use of external tools in data preparation, or specific customer deployments affect individual legal assessments.
As an affiliate, we earn on qualifying purchases.
Evaluation After the Release
The next step for prospective users is to review the technical report and model card, obtain the weights from Hugging Face, and run evaluations suited to their own language, hardware and safety requirements. Publicly available independent testing would help clarify how the model compares with alternatives beyond Aleph Alpha’s reported results.
Aleph Alpha has released the model and described its intended capabilities, but the supplied source does not identify a later product milestone, independent benchmark schedule or update date. Users should watch for further model documentation, evaluation results and deployment guidance as they become available, and treat performance and compliance as matters to verify for each application.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is Kolibri?
Kolibri is Aleph Alpha’s open-weight large language model designed for German and English. It has 78.1 billion total parameters, with about 3.46 billion active for each token, according to the company’s technical materials.
What license does Kolibri use?
Aleph Alpha released Kolibri’s weights and configuration files under Apache 2.0. The company says it retains rights to its training code and methods.
Can Kolibri run on a regular computer?
The supplied materials do not establish a minimum system specification for every deployment. They report about 78 GB of weights at 8-bit floating point and state that the full model must be in memory, indicating substantial hardware requirements.
Has Kolibri been independently shown to outperform similar models?
Not in the source material provided. Aleph Alpha reports that Kolibri scored above every compared model of its size in German and English, but independent confirmation is not established here.
What does Aleph Alpha mean by calling Kolibri sovereign?
The company points to development and training on infrastructure in Germany and Finland, and to customers’ ability to deploy the model under their own control. The model card also notes that some training data was prepared using tools from outside Europe, so the term does not mean every influence or component was European.
Source: hn
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
