Aleph Alpha Kolibri: open weights, not an outside benchmark

Aleph Alpha released Kolibri on 3 October 2026. It is an English and German mixture-of-experts model. The card says the weights and configuration files are published under the Apache 2.0 license, not the architecture or the training method. The Hugging Face repository is Aleph-Alpha/Kolibri-1, and that card carries the same release date.

The model card lists 78,103,074,560 total parameters and 3,457,573,120 active parameters per token. It labels those as 78B total and 3.46B active. The tech report, and the blog’s comparison table, say 78.1 billion total and 3.46 billion active. The blog also says “78B total parameters, 3B active.” The precise figures are the card’s integers, not that shorter line. The card states the knowledge cutoff as 18 June 2026 for both German and English, and says that only affects implicit knowledge.

The card says quality and serving efficiency are validated up to 1,048,576 tokens, with 262,144 native and 1,048,576 by extrapolation. It recommends at most 262,144 tokens for serving efficiency and for complex tasks. Reasoning effort can be set to none, low, medium, or high.

Aleph Alpha’s blog, dated 03/10/2026, says Kolibri was built for regulated work in public administration, industrials, and aerospace. The blog says teams built the model in Germany and trained it on infrastructure in Germany and Finland, under European and German law, and that it was built with the EU AI Act, the GPAI Code of Practice, and GDPR in mind. Those training-location and legal points are the company’s claims. These sources do not include a third-party audit of them. The blog’s comparison table says pre-training finished on 11 September 2026.

A non-public predecessor, Kolibri Origin, has 30.6 billion total parameters. It finished pre-training on 11 June 2026 and was not publicly released.

The blog says its scores come from Aleph Alpha’s own harnesses and, where applicable, the highest reasoning effort for each model. The model card says most benchmarks use eval-framework, and that TerminalBench and SWE-Bench use Harbor, with Kolibri at reasoning effort high. They are not an outside leaderboard. These sources do not show an independent lab rerunning them.

In an interview, open weight is not one thing. Separate total parameters from active parameters: 3,457,573,120 active per token, out of 78,103,074,560 total. Aleph Alpha says Kolibri can run on-premise without sending internal data to third-party inference services. That is the company’s claim, not a third-party finding, and it is not a claim that Kolibri beats a frontier API. Downloadable weights are also not the same thing as an independent lab rerunning the benchmarks.

https://aleph-alpha.com/en/blog/kolibri-has-landed-a-sovereign-open-weight-model/ https://huggingface.co/Aleph-Alpha/Kolibri-1 https://aleph-alpha.com/downloads/tech-report.pdf