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 use a mixture-of-experts architecture, and the company says its evaluations place it ahead of compared models of similar size in both languages.
Aleph Alpha released Kolibri on October 3, 2026, an open-weight large language model for German and English whose weights and configuration files are available under the Apache 2.0 license. The company says the model was trained from scratch on infrastructure in Germany and Finland, and describes it as a model intended to support deployment under European legal and data-governance requirements.
Kolibri has 78.1 billion parameters in total, but its mixture-of-experts (MoE) design activates about 3.46 billion parameters for each token, according to the model card and technical report cited in Tejas Kumar’s report. It has a stated native context window of 262,144 tokens, with testing reported up to 1,048,576 tokens. The model supports tool calling and four reasoning settings: none, low, medium and high.
Aleph Alpha says Kolibri was trained on about 24 trillion tokens, more than a fifth of them German, using 768 NVIDIA B200 GPUs. Its published knowledge cutoff is June 18, 2026. The weights are hosted on Hugging Face. Apache 2.0 applies to the weights and configuration files; according to the report, Aleph Alpha retains rights to its training code and methods.
In the company’s evaluation, Kolibri scored above every compared model of its size in German and English. That result is Aleph Alpha’s own evaluation claim; the source material does not provide enough detail here to independently assess the benchmark design, model selection or scoring. The technical report, model card and launch announcement are the cited bases for the reported specifications.
European Control and Model Access
Kolibri’s release gives organizations a model they can download and operate on their own infrastructure, rather than relying only on a hosted service. That may matter to public agencies, businesses and regulated institutions that need to keep sensitive data within their own systems or want greater control over deployment. Open weights make inspection and adaptation possible, though they do not by themselves settle questions about legal compliance, safety or performance.
Aleph Alpha uses “sovereign” to describe both how the model was built and the deployment options offered to customers. The company says its teams built the model in Germany, trained it in Germany and Finland, and operated under European and German law. This framing is relevant as European governments and companies consider how to use AI while managing data protection and dependence on external providers. It remains the company’s description of the model and its supply chain, not an independent certification.
The MoE design also illustrates a trade-off in model deployment: less computation per token does not mean a small memory footprint. The report says the full model must be held in memory even though only part is active for each token. Its stated 8-bit FP8 weights require about 78 GB, before accounting for other system needs. That can reduce inference computation relative to using all parameters for every token, but it does not make the model suitable for every server or device.
open source large language model German English
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Design Choices for German Text
Kolibri’s architecture routes each token through a limited group of specialist networks. According to the technical report as summarized by Kumar, the model has 50 layers, each with 384 routed experts and one shared expert; the router selects six of the 384 for each token. This structure is intended to provide access to a much larger set of learned parameters while limiting how many are used at once. The model still carries the memory demands of its full parameter set.
The model also uses a 128,000-token vocabulary built with an algorithm Aleph Alpha calls UniBPE. Kumar reports that it can represent some German compound words in fewer tokens than the tokenizer used in OpenAI’s GPT-4o and GPT-5. The company’s report says Kolibri’s tokenizer uses 11.2% fewer tokens for German text than GPT-5’s tokenizer in the comparison it conducted. That figure describes token counts in that test, not a general measure of model quality or a guarantee of faster or cheaper results in every use.
The “sovereign” label does not mean every element of the development process originated in Europe. The report says the model card identifies the use of Google’s Gemma 4 to rephrase English web text, Mistral-NeMo for German text, and Qwen3-32B to label data used in quality filters. Aleph Alpha also says it filtered training data for political bias and signed the European Union’s General-Purpose AI Code of Practice.
“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 its launch post
AI model deployment infrastructure Germany
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Benchmark and Deployment Questions
The available source material does not give the full benchmark tables or enough methodology to verify Aleph Alpha’s claim that Kolibri scores above all compared models of its size in both languages. The exact comparison set, evaluation prompts and margin of performance are not established here. Independent testing will be needed to determine how the model performs across tasks and against current alternatives.
Other practical details also need clarification for prospective users, including hardware requirements beyond the reported weight memory, serving costs, performance under real workloads, and the scope of customer support. The release materials described in the source also do not establish whether the model’s training data or training methods are fully reproducible. Open weights allow users to run the model, but do not provide unrestricted access to all parts of its development.
Aleph Alpha’s sovereignty language should be read narrowly: training infrastructure and deployment control are distinct from the origin of all tools and data used in development. The reported use of models from Google, Mistral and Qwen in parts of data preparation shows that the claim does not mean the process relied exclusively on European-developed components. The precise legal and operational implications for individual customers will depend on their deployment and compliance arrangements.
As an affiliate, we earn on qualifying purchases.
Independent Testing After Release
The immediate next step is for developers and organizations to download the weights from Hugging Face and test Kolibri against their own German- and English-language tasks. Those trials can establish whether the model’s claimed capabilities, long context window and tool calling meet particular needs, and whether available hardware can handle the full model.
Further scrutiny of Aleph Alpha’s technical report and model card, alongside independent benchmark results, should help clarify how its scores compare with other models and what the tokenizer gains mean in practical use. No additional release date, planned model update or independent evaluation is specified in the source material.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is Kolibri?
Kolibri is Aleph Alpha’s open-weight large language model for German and English. The company released it on October 3, 2026.
What does open-weight mean for Kolibri?
The model’s weights and configuration files are available under the Apache 2.0 license, allowing users to download and run them subject to the license. Aleph Alpha retains rights to its training code and methods, according to the cited report.
How large is the model, and how much is active at once?
Kolibri has 78.1 billion parameters in total, with about 3.46 billion active for each token through its mixture-of-experts architecture. The full model still needs to be held in memory.
Does Aleph Alpha’s evaluation prove Kolibri is better than similar models?
No. Aleph Alpha reports that Kolibri outscored every compared model of its size in German and English, but that is a company evaluation claim. The available material does not provide enough benchmark detail to independently verify it.
Does “sovereign” mean Kolibri used only European technology?
No. Aleph Alpha uses the term for the reported development and hosting arrangements and for customer deployment control. The model card also identifies non-European models used in parts of data preparation, including Google’s Gemma 4 and Qwen3-32B.
Source: hn
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
