Aleph Alpha put a European open-weight MoE on the public shelf. Kolibri is the ship, Apache 2.0 is the licence line, and Hugging Face is the pickup window.
The company posted the model as Aleph-Alpha/Kolibri-1 and published a tech report as arXiv:2610.31945. The accompanying blog, "Kolibri has landed: a sovereign open-weight model," frames the release for public administration, industry, and aerospace buyers that want English-German capacity under a permissive licence.
Aleph Alpha ships Kolibri-1 as Apache 2.0 open weights
Kolibri is an English-German mixture-of-experts transformer with 78.1B total parameters and 3.46B active parameters per token. The report states the model activates 4.4% of its parameters per token. Post-training covers supervised fine-tuning on curated German and English data plus reinforcement learning on more than 1.2M internally curated tasks spanning reasoning, tool use, instruction following, code, and retrieval.
Training ran across 24T tokens in pre-training, mid-training, and long-context extension, with German accounting for more than 20% of the data mix by bytes. Aleph Alpha says it complemented German datasets with more than 2T German tokens curated and synthesised through its own pipeline. The report positions Kolibri on a Pareto frontier of quality versus serving cost among the open models it evaluated in English and German. Those comparisons are the authors' evaluation set; this Dispatch leaves external leaderboard sweeps for a later Ground Truth if buyer benches appear.
Weights sit on Hugging Face under Apache 2.0, the tech report is on arXiv, and the narrative framing lives on the Aleph Alpha blog. Pin the Hugging Face revision before any pilot so card text and local files stay aligned.
Aleph Alpha's Model Factory, aimed at sovereign European buyers
Aleph Alpha shipped the weights and the report. The company describes both Kolibri and an earlier Kolibri Origin iteration as products of its Model Factory, linking data, experimentation, training, evaluation, and deployment through version-controlled workflows. The stated audience is sovereign and specialised deployment, with emphasis on German and on regulated domains such as public administration, industry, and aerospace.
Platform teams that already mirror Hugging Face weights, and European public-sector buyers writing open-weight procurement language, sit closest to the access matrix.
A 3.46B-active EN/DE MoE with a million-token card and FP8 weights
Relative to prior Aleph Alpha open artefacts, Kolibri's public card adds a concrete serving shape: FP8 weights, tool-use and reasoning post-training called out in the report, and a context window listed up to 1,048,576 tokens with a recommendation to stay at or below 262,144 for efficiency. Knowledge cutoff on the card for English and German: 18 June 2026. Card-dated release: 3 October 2026.
The Apache 2.0 line is the commercial signal. Buyers who needed a permissive European EN/DE MoE with a published active-parameter count now have a dated artifact to pin. Tool-use post-training in the report matters for agent pilots, though this piece stops short of claiming unstated tool-bench wins.
Card ceilings, author Pareto claims, and missing buyer benches
The Hugging Face card dates the release to 3 October 2026 and lists an English/German knowledge cutoff of 18 June 2026. Apache 2.0 covers the weights on the card and in the report. The focus is English-German with sovereign European positioning. Serving-cost claims in the report are relative Pareto statements on Aleph Alpha's evaluated open set; absolute dollars per million tokens remain undisclosed here. The card's 1,048,576-token maximum comes with an efficiency recommendation at 262,144.
Independent third-party board sweeps beyond the report's own tables are unsettled in this piece. Do not treat the blog's sovereign framing as a certified compliance pack for any named regulator.
Pin the HF revision and read Apache 2.0 against your procurement clause
Model platform owners should mirror Aleph-Alpha/Kolibri-1, pin the revision hash, and record the FP8 plus context recommendations before any pilot. Procurement and counsel in public admin or aerospace accounts should map Apache 2.0 against their open-weight clauses and export rules this month. Evaluation owners should run EN/DE task suites they already trust rather than copying the report's Pareto slide into a board deck.