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 German- and English-language model under the Apache 2.0 license. The company says its mixture-of-experts design activates about 3.46 billion of its 78.1 billion parameters per token; independent performance and deployment details remain limited in the supplied reporting.
Aleph Alpha released Kolibri on Oct. 3, 2026, as an open-weight large language model designed for German and English. The model’s weights and configuration files are offered under the Apache 2.0 license, while its mixture-of-experts design has 78.1 billion parameters in total and uses about 3.46 billion for each token, according to the company’s technical materials as described in a report by Tejas Kumar.
Kolibri is described as a mixture-of-experts (MoE) model: rather than activating every parameter for every token, a routing system selects a subset of specialized components. Kumar’s account of Aleph Alpha’s technical report says the model has 50 layers, each with 384 routed experts and one shared expert; each token is sent to six routed experts. The resulting active parameter count is reported as 3.46 billion, or about 4.4% of the total.
The model was trained from scratch on infrastructure in Germany and Finland, using 768 NVIDIA B200 GPUs and about 24 trillion tokens, according to the source material. More than a fifth of the training tokens were German. Kolibri’s reported context window is 262,144 tokens natively, with testing extending to 1,048,576 tokens. Its knowledge cutoff is listed as June 18, 2026, and it supports tool calling and four reasoning settings: none, low, medium and high.
Aleph Alpha says its evaluation places Kolibri above every compared model of similar size in German and English. That is a company-reported comparison; the material provided does not include the full benchmark tables, evaluation methods or results needed to independently assess the claim. The source also says the weights are available on Hugging Face. Apache 2.0 applies to the weights and configuration files, while Aleph Alpha retains rights to its training code and methods.
What Kolibri Offers European Deployers
Kolibri’s release gives organizations another model they can download and deploy under Apache 2.0, with a particular focus on German and English. For public agencies, manufacturers and other organizations handling sensitive information, the option to operate a model on their own infrastructure may matter as much as benchmark performance. Local operation can let customers keep prompts and documents within their own systems, though the practical level of control depends on deployment choices and the applicable contract and security arrangements.
The MoE architecture also illustrates a trade-off: Kolibri is reported to perform a smaller amount of computation per token than its total parameter count might suggest, but all model weights still need to be held in memory. Kumar reports that an 8-bit representation requires about 78 GB for the weights alone. That can affect hardware requirements and cost; activating fewer parameters does not make the full model equivalent to a 3.46-billion-parameter model to host.
For German-language applications, the tokenizer is another stated design focus. Aleph Alpha’s report says Kolibri’s tokenizer uses 11.2% fewer tokens for German text than GPT-5’s tokenizer in the company’s comparison of nine alternatives. Fewer tokens could affect processing efficiency for some workloads, but that figure is specific to the comparison described and does not establish lower total operating cost or stronger quality across tasks.
As an affiliate, we earn on qualifying purchases.
How Aleph Alpha Describes Sovereignty
Aleph Alpha uses the term “sovereign” for Kolibri, linking it both to how the model was developed and to customer deployment. In the company’s account, teams built the model in Germany and trained it on infrastructure in Germany and Finland, under European and German law. It says customers have freedom to deploy the model and can retain control over their intellectual property and data.
Those claims do not mean the model was developed without tools or materials originating outside Europe. The model card, as summarized by Kumar, 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. Aleph Alpha also says it filtered training data for political bias. These details distinguish the location and governance claims from a claim that every component or source was European.
The company has signed the European Union’s General-Purpose AI Code of Practice, according to the source. Kumar’s report draws on Aleph Alpha’s 189-page technical report, model card and launch post, alongside his own tokenizer experiment. The information here reflects those materials as recounted in the supplied report, rather than a separate independent audit of training, benchmarks or compliance.
“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 launch materials described by Tejas Kumar
large language model deployment hardware
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Benchmark and Hosting Questions Remain
The supplied report does not provide enough benchmark detail to verify Aleph Alpha’s claim that Kolibri scores above every compared model of its size in both languages. The models compared, task selection, scoring methods and independent replication are not established in the material provided. The claimed tokenizer reduction is also a company-reported measurement and should not be treated as a general performance or cost guarantee.
It is not yet clear from the supplied information how Kolibri performs across specific real-world tasks, how it compares with other available models on safety and reliability, or what deployment configurations Aleph Alpha recommends. The reported 78 GB of 8-bit weights does not include all possible system requirements, such as additional memory for inference and longer contexts. Nor does the company’s sovereignty description by itself establish the legal or technical terms of every customer deployment.
The source describes a tokenizer experiment by Kumar but cuts off before giving its results. It also does not provide detailed findings on model behavior, the full content of the technical report’s evaluations, or independent confirmation of the training and data-governance claims. Those points remain open on the basis of the material available here.
As an affiliate, we earn on qualifying purchases.
Independent Tests and Deployment Details
The immediate next step for prospective users is to review the model weights, configuration files and model card on Hugging Face, then test the model on their own German- and English-language tasks. Organizations will need to check hardware needs, license terms and operating requirements before choosing a deployment. The reported context limits and per-token active parameter count do not by themselves establish how the model will perform in a particular environment.
Further assessments could clarify the company’s benchmark comparisons, the model’s reliability across use cases, and the practical costs of serving it. Independent testing of the reported tokenizer efficiency and the model’s long-context behavior would also help users judge where Kolibri fits. Aleph Alpha’s detailed technical report is the cited source for these evaluations, but the materials supplied do not specify a date for additional results or a future release.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is Kolibri?
Kolibri is an open-weight large language model from Aleph Alpha, designed for German and English. It is described as a mixture-of-experts model with 78.1 billion total parameters.
When was Kolibri released, and what license does it use?
Aleph Alpha released Kolibri on Oct. 3, 2026. Its weights and configuration files are offered under the Apache 2.0 license; the company retains rights to its training code and methods.
How many parameters does Kolibri use for each token?
The model has 78.1 billion parameters in total and is reported to activate about 3.46 billion parameters per token. Its full weights still need to be held in memory, according to the model card quoted in the source report.
Does Aleph Alpha’s performance claim have independent confirmation?
The company says Kolibri scored above every compared model of its size in German and English. The supplied material does not include enough benchmark detail or independent results to verify that claim.
What does Aleph Alpha mean by calling Kolibri sovereign?
Aleph Alpha points to development and training in Germany and Finland under European and German law, along with customer deployment control. The model card also identifies some non-European tools used in preparing training data, so “sovereign” does not mean every input or component originated in Europe.
Source: hn
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
