TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Aleph Alpha released Kolibri on Oct. 3, 2026, as an open-weight model for German and English under the Apache 2.0 license. It has 78.1 billion total parameters but activates about 3.46 billion per token; the company reports strong results against selected models, though independent verification and practical deployment details remain limited in the available material.
Aleph Alpha released Kolibri on Oct. 3, 2026, an open-weight language model for German and English whose weights and configuration files are available under the Apache 2.0 license. The model uses a mixture-of-experts design with 78.1 billion parameters in total, while activating about 3.46 billion for each token, a design intended to reduce computation without reducing the model’s full stored size.
Aleph Alpha’s technical report and model card describe Kolibri as trained from scratch on approximately 24 trillion tokens, with more than one-fifth of its training data in German. The report says training used 768 NVIDIA B200 GPUs on infrastructure in Germany and Finland. The model has a native context window of 262,144 tokens; the supplied source says it was tested up to 1,048,576 tokens, but does not give detailed results for that longer context.
Kolibri’s mixture-of-experts architecture contains 50 layers, each with 384 routed experts and one shared expert. For each token, its router selects six of the 384 routed experts, alongside the shared component. This means fewer parameters are active during each computation than the total stored in the model. However, according to the model card, the full model must still be held in memory. The source estimates about 78 GB for weights in FP8 format, so lower per-token computation does not make Kolibri a small-memory model.
Aleph Alpha says its evaluations put Kolibri above every compared model of similar size in both German and English. That is a company-reported evaluation; the provided material does not include enough benchmark tables or independent testing to establish how it performs across different tasks. The model supports tool calling and four reasoning settings—none, low, medium and high—and its reported knowledge cutoff is June 18, 2026.
A European Option for Model Deployment
Kolibri’s release gives organizations a model they can download and deploy under an established open-source license, rather than relying only on a hosted service. That can matter to public agencies, companies and research teams that want to keep prompts and internal data on their own systems. Aleph Alpha describes the model as sovereign and says it was built and trained under German and European law, with no foreign control over the training infrastructure.
Those descriptions are relevant to buyers weighing data handling, legal accountability and control over deployment, but they should not be read as proof that every use of the model meets a particular organization’s compliance requirements. Deployment, data governance and risk assessments still depend on the customer’s setup. The architecture also involves a practical trade-off: reduced active computation, but substantial memory needs to load all experts.
The release also targets a language-specific technical issue. German compound words can be split into more tokens by tokenizers optimized for other languages. Aleph Alpha says Kolibri’s tokenizer uses 11.2% fewer tokens for German text than the GPT-5 tokenizer in its comparison of nine alternatives. Fewer tokens may affect processing costs and available context, but the reported figure is a vendor measurement, not an independent estimate of savings for every workload.
As an affiliate, we earn on qualifying purchases.
Training, Licensing and Sovereignty Claims
The model’s weights are hosted on Hugging Face, and Aleph Alpha released the weights and configuration files under Apache 2.0. The company retains rights to its training code and methods, according to the supplied report. The license therefore applies to the published model materials identified by the source, not necessarily every component of the development process.
Aleph Alpha says the model was developed with the EU AI Act in mind and has signed the European Union’s General-Purpose AI Code of Practice. Its account of sovereignty focuses on European teams, infrastructure and legal jurisdiction, as well as customers’ ability to deploy the model themselves. That framing does not mean every element of the training pipeline originated in Europe: the model card says English web text was rephrased using Google’s Gemma 4, German text using Mistral-NeMo, and Qwen3-32B was used to label data for quality filters.
The release date and technical figures in this article come from the supplied report, which says its author drew on Aleph Alpha’s 189-page technical report, model card and launch post, plus an experiment with the tokenizer. The source includes the author’s own positive assessment of the project; that opinion is distinct from the reported specifications and company evaluation claims.
“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, as quoted in the supplied report
large language model deployment hardware
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Independent Testing and Use-Case Results
The supplied material does not provide independent benchmark results or enough detail to assess the company’s comparisons across tasks, prompting methods and hardware. Kolibri’s performance outside Aleph Alpha’s evaluations, including how it compares on specialized German-language tasks, remains to be established through broader testing.
It is also unclear from the available source how much compute and memory are required for particular deployment configurations, what performance users can expect at the tested one-million-token context length, and how the model behaves in extended real-world use. The published license covers the weights and configuration files, while the precise practical scope of rights and restrictions across other training materials is not detailed here.
The model card’s account of third-party models used to rephrase or label training data also leaves open questions about the quantities and influence of that data. The source says Aleph Alpha filtered training data for political bias, but provides no specific measurement in the supplied material that would allow readers to judge the effectiveness of that work.
open-source AI model training tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Testing After the Release
The next useful evidence will come from developers and organizations running Kolibri on their own hardware and publishing results. Comparisons should identify the task, language, model version, hardware and evaluation method, so readers can distinguish general claims from performance in a specific deployment.
Potential users can inspect the weights, configuration and model card on Hugging Face, then assess whether their systems can support the model’s memory demands and whether its license and data practices suit their needs. Aleph Alpha’s release announcement and technical documentation are the immediate sources for specifications; further documentation or independent evaluations may clarify performance, costs and long-context behavior.
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. The company released its weights and configuration files under the Apache 2.0 license on Oct. 3, 2026.
How large is Kolibri?
It has 78.1 billion parameters in total and activates about 3.46 billion for each token through its mixture-of-experts design. The full model still needs to be held in memory, according to its model card.
Can organizations run Kolibri themselves?
The weights and configuration files are published for download, so organizations can deploy the model themselves, subject to the license and their technical setup. The supplied source estimates about 78 GB of weights in FP8; actual deployment requirements can vary.
Has Kolibri’s performance been independently confirmed?
The supplied report says Aleph Alpha’s evaluations placed Kolibri above every model it compared with at a similar size in German and English. That is a company-reported result; independent results are not included in the source material.
What does Aleph Alpha mean by calling the model sovereign?
The company points to European development and training infrastructure, European and German legal jurisdiction, and customers’ ability to deploy the model themselves. The model card also reports use of some third-party models in preparing training data, so sovereignty here does not mean every pipeline component was European-made.
Source: hn
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
