Aleph Alpha Releases Open-Weights Model Kolibri: 78B Total Parameters, 3B Active, 21.3% German Training Data
The Decoder reported on October 5, 2026 that Aleph Alpha released Kolibri, a German-English bilingual MoE model with 78 billion total parameters, roughly 3 billion active per token, German making up 21.3% of training data, and weights released under Apache 2.0 on Hugging Face, supporting up to one million tokens of context.
What changed
The Decoder reported on October 5, 2026 that German company Aleph Alpha released Kolibri, a German-English bilingual language model. The model uses a mixture-of-experts architecture with 78 billion total parameters, activating only about 3 billion parameters per token, and offers a context window of up to one million tokens. According to the report, German accounted for 21.3% of its training data, backed by a German-language data pipeline the company built specifically for the project; Chinese models were also used in synthetic training data, though the report did not identify which ones. Aleph Alpha said Kolibri scored 71% on German benchmarks while decoding faster than comparable models listed in the report, including GPT OSS A5B, Qwen 3.6 A3B, and Gemma 4 A4B, and claimed it sits on the Pareto frontier of quality and operating cost in both German and English. According to the technical report, the model was trained on 768 B200 GPUs in Germany and Finland, with the development process taking the EU AI Act into account and positioning it for public administration, aviation, and industrial use cases. The weights have been released on Hugging Face under the Apache 2.0 license. It is important to distinguish open weights from an open-source license; here the report explicitly states the specific Apache 2.0 license, but the actual scope of those license terms should still be governed by the model card and the license text itself.
What it means for developers
The direct implications for developers center on three points. First, an MoE structure with roughly 3 billion active parameters alongside 78 billion total parameters means inference compute is closer to that of a small model while memory footprint is closer to that of a large one, so the deployment cost profile differs from a dense model of the same size, and implementation details of quantization and expert routing will significantly affect throughput. Second, the one-million-token context window is a long-context capability claim, but the report gives no retrieval accuracy, latency, or memory data at that length, so long-document scenarios need independent validation and cannot simply inherit benchmark conclusions from shorter contexts. Third, the 21.3% German share of training data and the dedicated German data pipeline suggest the model should perform better on German tasks than most English-dominant open-weights models, but the reported 71% German benchmark score is a vendor claim that does not specify the benchmark name, evaluation version, or whether it was zero-shot, so cross-model comparisons should be cautious. In addition, Apache 2.0 is a relatively permissive license, which is important information for teams hoping to use the model commercially or self-host it, but the license does not cover the compliance of training data, and responsibility related to the EU AI Act still rests with the deployer. The report mentions that Chinese models were used to generate synthetic data during training; this is a disclosed fact and should not be read as Kolibri having any technical lineage relationship with any Chinese model.
Integration and evaluation
The following is advisory analysis based on the report, not verified fact. The first decision is model selection: if the workload is primarily German and requires self-hosting, Kolibri can be listed as a candidate; if it is primarily Chinese or English, its German data pipeline advantage does not constitute a reason for selection. A suggested evaluation condition is to build a set of German and English tasks covering the target industry, with a quantitative criterion of comparing German-task accuracy and end-to-end latency per thousand tokens under the same hardware and same quantization configuration, rather than relying only on the single 71% figure in the report. The second decision is the context-versus-cost tradeoff: the report claims support for up to one million tokens but gives no performance data at that length, so a suggested approach is to establish a baseline at moderate lengths of 8K to 32K and then gradually stress-test up to the target length, with evaluation criteria being whether key-information recall degrades significantly with length and whether first-token latency falls within an acceptable range. The third decision is the licensing and compliance path: although the weights are released under Apache 2.0, the model is positioned for regulated industries such as public administration and aviation, so deployers should confirm the model card's specific statements on redistribution, fine-tuned derivatives, and liability allocation, with the evaluation criterion being whether legal review can find clear clauses in the license text and model card covering their own use case.
Limits and open questions
First, it must be made clear that as of this catalog check, the Kolibri version does not appear in KimiSeek's live model catalog; the weight distribution and any API availability discussed in the report point to Aleph Alpha and its Hugging Face page and are unrelated to KimiSeek. Second, statements in the report about the Pareto frontier, faster decoding than comparable models, and the 71% German benchmark score all come from Aleph Alpha's own claims; this article relays them rather than independently verifying them, and the report does not specify benchmark names, evaluation methods, whether results were zero-shot, or the configurations of competing models. Third, the specific versions and release dates of the comparison models GPT OSS A5B, Qwen 3.6 A3B, and Gemma 4 A4B are outside the scope of this check, so their comparability cannot be confirmed. Fourth, Apache 2.0 governs the distribution license for weights and code and does not constitute a compliance endorsement of training-data provenance, and the report's detail about using Chinese models for synthetic data does not identify the specific models or generation methods. Fifth, the training scale of 768 B200 GPUs and the German-English bilingual positioning are vendor descriptions; reproduction costs, the amount of data needed for fine-tuning, and true performance in languages other than German remain unknown.
Illustrations from the original report
