Aleph Alpha releases Kolibri, an open-weight 78B model for regulated sectors

What Kolibri is and how it can be used

Kolibri supports context lengths of up to 1M tokens. The full weights can be downloaded on Hugging Face and used under Apache 2.0 terms. Aleph Alpha published the model on 3 October 2026, which it describes as the Day of German Reunification, and refers readers to a tech report for full details.

The company says the model's small and efficient size lets customers run it on-premise, without sending internal data to third-party inference services. It also says Kolibri supports controllable reasoning effort, so users can trade cost and latency against answer quality.

Kolibri was specialized for German, reasoning, math, agentic behavior and other capabilities that customers need in production. Aleph Alpha says the goal was to optimize performance for its customers' specific use cases and to let them monitor the economic impact, so that return on investment stays measurable.

Benchmark results and the efficiency claim

Aleph Alpha says Kolibri sits on the Pareto frontier for quality versus serving cost, in both English and German. In the company's words, none of the compared models delivers more quality at the same serving cost, or the same quality at lower cost. It also says that across math, coding, grounding and long-context tasks, Kolibri matches models with up to four times its active parameter count, such as Nemotron 3 Super.

The company compared Kolibri with its predecessor Kolibri Origin, Qwen3.6-35B-A3B, Nemotron 3 Super 120B-A12B and Mistral Small 4 119B-A6B. Kolibri scored 96.9 on AIME 2025, against 91.7 for Nemotron 3 Super and 84.6 for Qwen3.6-35B-A3B. On the German version of that test it scored 87.5. It reached 84.3 on GPQA (diamond), 85.9 on LiveCodeBench v6 and 29.4 on BrowseComp.

The results are not uniformly in Kolibri's favor. Qwen3.6-35B-A3B scored higher on BFCL v4 overall (67.2 versus 61.4), on τ²-bench telecom (99.1 versus 94.7) and on LongBench Pro (70.8 versus 64.5). On the AA-Omniscience Index, Kolibri's score of -32.8 was behind Qwen3.6-35B-A3B at -15.3 and Mistral Small 4 at -24.0. Aleph Alpha describes Kolibri as a balanced generalist with competitive performance across math, code, long context, agentic capabilities and knowledge.

Specialization, grounding and German language focus

Because public benchmarks do not capture specialized sector needs, Aleph Alpha built internal evaluation suites for verticals including the German public sector, aviation, manufacturing and the automotive industry. Paired synthetic training environments let the company improve the model against these evaluations, it says, without training on customer data.

On these internal customer-proxy benchmarks, scores rose across successive post-training runs in all five verticals. The automotive supplier suite went from 0.72 to 0.99 and semiconductors from 0.35 to 0.80. The German public sector suite moved from 0.54 to 0.75, industrial drive technology from 0.31 to 0.60 and aerospace from 0.14 to 0.59.

Kolibri was also trained with abstention data and with the company's Merlin-Arthur protocol. As a result, according to Aleph Alpha, it is trained to say "I don't know" when the answer is not in the context. The company says customers request this feature and that it tracks abstention accuracy continuously.

For language, the company built a bilingual German/English tokenizer. German makes up 21.3% of the pre-training tokens, and translation was used sparingly, at 6% overall, because translated text tends to carry the cultural fingerprint of its source language. Aleph Alpha calls the result a model that is bilingual by design.

Sovereignty claims and the Model Factory

Aleph Alpha says it built Kolibri with the EU AI Act, the General-Purpose AI Code of Practice and the GDPR in mind, with copyright law a particular focus. The company says it is transparent about model weights and training data curation, and that reasoning traces make it explainable how the model reached an answer.

The team built the model in Germany and trained it on infrastructure in Germany and Finland, under European and German law, with "no foreign control", according to the company. It says it owns the entire pipeline, from data curation through pre- and post-training to optimization in its Model Factory, and that customers get full freedom of deployment and intellectual-property safety.

The Model Factory is the company's training pipeline implemented as code, so that the team's learnings sit in one versioned training recipe rather than scattered scripts and notes. Work on it began in January. Kolibri Origin, a 30B total, 3B active model with a 65k-token context window, finished pre-training on 11 June after hundreds of ablation runs. Kolibri finished on 11 September. In those three months the company went from 30B to 78B parameters, from a 65k-token context window to up to 1M, and from 7.5T to 20T training tokens. Aleph Alpha says pre-training ran without a person having to step in when hardware failed or a data connection dropped.

What it means for users

For organizations in regulated sectors, Kolibri offers open weights under a permissive license and a model small enough, by the company's account, to run on-premise. Aleph Alpha's efficiency and benchmark claims come from its own evaluations, and the internal customer-proxy results are measured on suites the company built itself. The tech report is the place to check the details.