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 large language model designed for German and English. Its Apache 2.0-licensed weights are available on Hugging Face; Aleph Alpha reports a 262,144-token native context window and says Kolibri outperformed the models it compared in its size range. Independent performance results and several deployment details remain unverified in the supplied source.
Aleph Alpha released Kolibri, an open-weight large language model for German and English, on October 3, 2026, publishing its weights on Hugging Face under the Apache 2.0 license. The model uses a mixture-of-experts design with 78.1 billion total parameters, while activating about 3.46 billion for each token, a setup intended to limit computation while retaining a much larger model’s capacity.
Aleph Alpha’s technical report lists a 262,144-token native context window, four reasoning levels—none, low, medium and high—and support for tool calling. The company says Kolibri was trained from scratch on infrastructure in Germany and Finland, using 768 NVIDIA B200 GPUs and about 24 trillion tokens. More than one-fifth of the training tokens were German, according to the report. Its stated knowledge cutoff is June 18, 2026.
The model’s mixture-of-experts architecture has 50 layers, each with 384 routed experts and a shared expert. For each token, the router selects six of the 384 routed experts. That reduces the parameters active in processing each token, but does not reduce the full model’s memory needs: all its weights must still be available. Aleph Alpha estimates about 78 GB of memory for FP8 weights; actual deployment requirements can depend on hardware and software configuration.
The release includes model weights and configuration under Apache 2.0. The source report says Aleph Alpha retains rights to its training code and methods. The license therefore applies to the released model materials, not necessarily every component of the development process. The report also describes a tokenizer with a 128,000-token vocabulary, designed to handle German compounds more efficiently than some alternatives.
A European Option for German Workloads
Kolibri adds an openly licensed model built with German-language use and European deployment conditions in mind. For public agencies, companies and researchers, access to downloadable weights may offer more control over where a model runs and what data is sent to an outside service. That can matter for organizations handling sensitive records or operating under data-protection and sector rules. Whether the model meets a particular organization’s legal or security needs still depends on its deployment and use.
Aleph Alpha describes the model as sovereign, pointing to development and training in Germany and Finland and operation under European and German law. The company says customers can deploy it themselves and retain control over their data and model access. Those are company claims about the release and its intended deployment; they do not mean the model contains no contributions from outside Europe. The model card, as summarized in the source report, says external models were used to rephrase some training text and label data used in quality filtering.
The sparse architecture also makes the release relevant to organizations weighing performance against inference costs. Activating about 3.46 billion parameters per token may reduce computation compared with running all 78.1 billion at once. However, users still need to hold the full model in memory, so the active parameter count alone does not describe the hardware required. Benchmark results, operational costs and quality on specific tasks will matter more to prospective users than the headline parameter figures.
As an affiliate, we earn on qualifying purchases.
How Kolibri Was Built
The release follows Aleph Alpha’s work on language models and its stated focus on enterprise and public-sector use. The supplied report says Kolibri was trained from scratch rather than being presented as a fine-tune of a single existing model. Its technical report, model card and launch post are the cited materials behind the release details and company performance claims.
Aleph Alpha’s approach to German text includes a tokenizer intended to split long compound words into fewer pieces. The source report gives “Bundesverfassungsgericht” as an example: it says Kolibri represents the word in two tokens, compared with six under the OpenAI tokenizer it names. Aleph Alpha’s report says its tokenizer used 11.2% fewer tokens on German text than GPT-5’s tokenizer, the best result among nine alternatives it measured. This is a result from the company’s stated comparison, not an independent benchmark presented in the supplied material.
The company also says it considered the EU AI Act during development and has signed the European Union’s General-Purpose AI Code of Practice. The source report notes that training data included text processed with Google’s Gemma 4 and Mistral-NeMo, as well as quality-filter labels from Qwen3-32B. These details qualify the sovereignty claim: the reported location and legal setting of development do not mean every tool or model involved originated in Europe.
large language model deployment hardware
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Independent Tests Still Needed
The source material reports that Kolibri scored above every compared model of its size in both German and English, but the supplied summary does not give the benchmark names, scores, test settings or comparison details. The claim is therefore attributable to Aleph Alpha’s evaluation; readers cannot use this information alone to judge how Kolibri performs against other models on their own tasks.
It is also not clear from the supplied material how much memory or what hardware is required for different deployment configurations beyond the estimate for FP8 weights. The report mentions a context length tested up to 1,048,576 tokens, but gives 262,144 tokens as the native context window; the conditions and performance at the longer length are not detailed here. The full training data, complete evaluation methodology and practical operating costs are also not established by the supplied summary.
Finally, the available information does not independently verify the company’s claims about sovereign deployment, legal compliance or the model’s behavior across real-world use cases. The open license provides access to weights and configuration, but organizations will need to review the full license, documentation and applicable requirements before adopting the model.
AI model memory optimization tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
What Users Can Evaluate Now
Prospective users can download the weights and configuration from Hugging Face and review Aleph Alpha’s technical report and model card. The next practical step is testing the model on the intended languages, tasks and infrastructure, including checking response quality, latency, memory use, tool calling and behavior at longer context lengths.
Independent benchmark results and deployment reports would help clarify how Kolibri compares with other open-weight models beyond Aleph Alpha’s own evaluation. Organizations considering use in regulated or sensitive settings will also need to assess data handling, model limitations, licensing and compliance for their specific deployment. No additional release date or independent evaluation milestone is specified in the supplied material.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is Kolibri?
Kolibri is an open-weight LLM released by Aleph Alpha for German and English. The company describes it as a mixture-of-experts model with 78.1 billion total parameters and about 3.46 billion active per token.
When was Kolibri released, and what is its license?
Aleph Alpha released Kolibri on October 3, 2026. The weights and configuration are available under the Apache 2.0 license; the source report says the company retains rights to its training code and methods.
Does Kolibri need only 3.46 billion parameters’ worth of memory?
No. That figure describes the parameters active for each token. The full model must still be held in memory, and Aleph Alpha estimates about 78 GB for FP8 weights.
Has Kolibri’s performance been independently confirmed?
The source report attributes the claim that Kolibri scored above compared models of its size in German and English to Aleph Alpha’s evaluation. The supplied material does not provide independent benchmark results or enough test details to verify that comparison.
Source: hn
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
