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 October 3, 2026, as an open-weight model for German and English, with its weights and configuration files under the Apache 2.0 license. The company describes it as a European-built mixture-of-experts model, but its performance claims come from its own evaluations and independent results are not provided in the source material.
Aleph Alpha released Kolibri, an open-weight large language model for German and English, on October 3, 2026. The model’s weights and configuration files are available under the Apache 2.0 license, while the company presents its European development and hosting arrangements as a way for organizations to retain control over deployment and data.
Kolibri is a mixture-of-experts model with 78.1 billion parameters in total, according to Aleph Alpha’s technical report and model card as summarized by Tejas Kumar. The company says only about 3.46 billion parameters are active for each token. That can reduce the computation required per token compared with a dense model of the same total size, but it does not remove the need to hold the full model in memory.
The model is reported to have a 262,144-token native context window, with testing reported up to 1,048,576 tokens. Its training used about 24 trillion tokens and 768 NVIDIA B200 GPUs, with more than a fifth of the training tokens in German, according to the supplied report. The model card lists June 18, 2026, as its knowledge cutoff and describes four reasoning settings: none, low, medium and high.
Aleph Alpha says Kolibri was trained from scratch on infrastructure in Germany and Finland, under European and German law. The company’s use of “sovereign” refers to development and deployment control; it does not mean every component of the process originated in Europe. The model card, as described by Kumar, says Google Gemma 4 was used to rephrase English web text, Mistral-NeMo for German text and Qwen3-32B for labeling data used in quality filters.
German Language and Deployment Control
Kolibri combines two selling points for organizations that use German-language AI: a model designed with German text in mind and the option to run its weights on infrastructure they control. That may appeal to public agencies, manufacturers and other organizations with strict data-handling requirements, particularly where sending information to an external service is not acceptable. The source material does not establish how widely customers will deploy the model or whether it meets any particular organization’s compliance needs.
The model’s sparse design also illustrates a practical trade-off. Although only a small share of its parameters is active for each token, the full set of weights must still be available in memory. Kumar reports that an 8-bit representation requires about 78 GB for the weights; that figure describes weight storage, not the full hardware requirements for running a deployment. Organizations will need to weigh that memory demand against the model’s potential computational savings and the costs of operating it.
Aleph Alpha says its evaluation placed Kolibri above every model of comparable size that it tested in both languages. That is a company-reported result, not an independently verified ranking in the supplied material. Its significance will depend on the benchmarks, comparison models and task performance, as well as how well the model works in real deployments.
As an affiliate, we earn on qualifying purchases.
How Kolibri Is Built
Kolibri uses a mixture-of-experts, or MoE, architecture. In a dense model, each token passes through the same broad set of parameters. In an MoE model, a routing system selects a subset of specialist networks for each token. Kumar’s account of the technical report says Kolibri has 50 layers, each with 384 experts and one shared expert; the router sends each token to six of the 384 experts. This structure explains why its total parameter count is much larger than its reported active count.
The model also has a 128,000-token vocabulary and a tokenizer Aleph Alpha calls UniBPE. The company designed it to account for German compounds, which can be split into many fragments by tokenizers less suited to German. Kumar gives “Bundesverfassungsgericht” as an example: Kolibri’s tokenizer represents it in two tokens, while the OpenAI tokenizer he tested represents it in six. The report says Kolibri’s tokenizer used 11.2% fewer tokens on German text than GPT-5’s tokenizer in its comparison of nine alternatives; this is a reported tokenizer result, not a measure of overall model quality.
Aleph Alpha also says Kolibri was developed with the EU AI Act in mind and that it has signed the EU’s General-Purpose AI Code of Practice. These points describe the company’s approach and commitments; they do not by themselves establish regulatory approval or certify that every deployment satisfies legal requirements.
“The full model must be held in memory even though only part of it is active at any time.”
— Aleph Alpha, as quoted in the model card and summarized by Tejas Kumar
large language model deployment hardware
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Independent Results and Deployment Needs
The supplied material does not include independent benchmark results or enough detail to assess Aleph Alpha’s comparison with models of similar size. It is also unclear how Kolibri performs across different German dialects, specialist tasks, factuality tests and real-world workloads beyond the company’s reported evaluations.
Actual deployment requirements will depend on the format, precision, serving setup and workload. The approximately 78 GB weight estimate at 8-bit precision does not specify total system memory, speed or operating costs. The source also does not provide independently checked energy-use figures, customer deployments or evidence that a particular use of the model complies with applicable law.
As an affiliate, we earn on qualifying purchases.
Benchmarking and Adoption Tests
The weights and configuration files are available through Hugging Face, giving developers a route to inspect and test the model. The next useful evidence will come from evaluations that disclose their methods and comparison baselines, alongside reports from organizations running Kolibri on their own infrastructure. Those tests can show whether its German-language design, long context and active-parameter efficiency translate into practical advantages.
Until such results are available, readers should treat the reported benchmark and tokenizer comparisons as Aleph Alpha’s claims, and distinguish the model’s open-weight license from the separate rights the company retains over its training code and methods.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is Kolibri?
Kolibri is Aleph Alpha’s open-weight language model for German and English. The company released it on October 3, 2026.
What does open-weight mean for Kolibri?
Its weights and configuration files are released under Apache 2.0, according to the supplied report. Aleph Alpha retains rights to its training code and methods, so open weights do not mean every part of the development process is open.
How large is the model?
Kolibri has 78.1 billion parameters in total, with about 3.46 billion active for each token. The full model still needs to be held in memory.
Is Kolibri independently proven to outperform comparable models?
No independent confirmation is included in the source material. Aleph Alpha says its evaluations put Kolibri ahead of every model of comparable size that it tested in German and English; the basis and results should be checked against the company’s technical report.
Does “sovereign” mean Kolibri uses only European technology?
No. Aleph Alpha uses the term for its stated development and deployment control in Europe. The model card also reports using Google Gemma 4, Mistral-NeMo and Qwen3-32B in parts of the data preparation process.
Source: hn
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
