Aleph Alpha releases Kolibri, a 78B-parameter open-weight model for sovereign AI

What Kolibri is

Aleph Alpha has released Kolibri, an open-weight language model aimed at what the German company calls sovereign, mission-critical work in regulated fields such as public administration, industry and aerospace. The announcement on Aleph Alpha's blog, which also reached the Hacker News front page, was timed to the Day of German Reunification on 3 October 2026.

Technically, Kolibri is a Mixture-of-Experts transformer with 78 billion total parameters, of which roughly 3 billion are active per token. It is bilingual in English and German and supports context lengths of up to 1 million tokens. The full weights are downloadable from Hugging Face under Apache 2.0 terms, so organizations can run, fine-tune and deploy the model without a separate commercial licence or reliance on a hosted inference API.

A pipeline proven on Kolibri Origin

According to Aleph Alpha, Kolibri is the second model through a training pipeline the company built end to end, covering data ingestion and curation, ablations, pre-training, post-training and final evaluations. The pipeline was first validated with Kolibri Origin, an earlier 30B-total / 3B-active model with a much shorter 65,000-token context window.

The company says the infrastructure allowed hundreds of ablation experiments and pre-training runs stable enough to recover automatically from hardware failures and dropped data connections, and that the short gap between the two releases reflects that investment. A full technical report accompanies the release.

Self-reported benchmarks

Aleph Alpha's own benchmark table compares Kolibri against Qwen3.6-35B-A3B, Nemotron 3 Super 120B-A12B and Mistral Small 4 119B-A6B. Reported results include 96.9 on AIME 2025, 90.0 on the German-language AIME 2026 variant, 84.3 on GPQA diamond, 85.9 on LiveCodeBench v6 and 92.7 on HumanEval+. The company claims Kolibri matches models with up to four times its active parameter count, such as Nemotron 3 Super, across math, coding, grounding and long-context tasks, and that it sits on the Pareto frontier of quality versus serving cost in both English and German — that is, no compared model in its tests delivered more quality at the same serving cost, or the same quality at lower cost. These are vendor-run figures with no independent verification cited.

Beyond public benchmarks, Aleph Alpha built internal "customer-proxy" evaluation suites for five verticals and reports large gains from Kolibri Origin to Kolibri: automotive supplier 0.72 to 0.99, semiconductors 0.35 to 0.80, German public sector 0.54 to 0.75, industrial drive technology 0.31 to 0.60, and aerospace 0.14 to 0.59. The company says paired synthetic training environments let it improve against these suites without ever training on customer data.

Grounding, German data and sovereignty

Two design choices stand out for regulated customers. First, Kolibri was trained with abstention data and with Aleph Alpha's Merlin-Arthur protocol, so it is explicitly taught to say "I don't know" when an answer is not present in the provided context. The company says it continuously tracks abstention accuracy, and publishes a hallucination-focused AA-Omniscience Index on which Kolibri scores -32.8, behind the compared Qwen model at -15.3.

Second, the model is German by construction rather than by afterthought: a custom German/English tokenizer, organic German text making up 21.3% of pre-training tokens, and translation used sparingly at 6% overall, since translated text tends to carry the cultural fingerprint of its source language.

Sovereignty, in Aleph Alpha's framing, covers both how the model was built and how it transfers: full supply-chain accounting from data ingestion to evaluation, transparency into every training decision, freedom of deployment, and intellectual-property safety, with compliance arriving as an inherited property. The small active parameter count keeps serving costs low enough for efficient on-premise deployment, so internal data never has to leave the customer's environment.

Why it matters

Kolibri is a notable data point in the European AI debate. Governments and regulated industries have repeatedly expressed concern about depending on US-hosted frontier models, and Aleph Alpha is now offering a concrete alternative: genuinely open Apache 2.0 weights, a 1M-token context window, strong German-language performance and grounding behaviour tuned for administration and industry — a combination rarely found in one package.

The sparse Mixture-of-Experts design, with only 3B active parameters, is also part of a broader industry shift toward cheaper inference, and Aleph Alpha claims it beats larger models on cost-quality trade-offs. The caveat is that every performance claim here comes from the vendor's own evaluations. Independent benchmarking will determine whether Kolibri is competitive in practice, but as a sovereignty-focused, fully downloadable model from a prominent European lab, the release itself moves the European open-model ecosystem forward.