Aleph Alpha

Overview

Aleph Alpha is a German AI lab. This page was created on 2026-10-04 on the release of Kolibri-1, an Apache-2.0 open-weight Mixture-of-Experts model, and its contents are limited to what that release and the lab's own blog index stated (source).

The lab's founding date, funding, headcount, ownership and its earlier model lines were not established on the run that created this page. aleph-alpha.com answered normally and its blog index was read, but that index carries no company page and no dates against its posts, and no other primary page was fetched. So what is here is one release and a reading of the lab's published writing — not a company profile.

What that writing does establish is a consistent position: the blog argues specifically for German-language training data and for sovereign deployment, and the Kolibri release is the artefact matching it.

Key People

unknown — no named individual appears in any source read on the run that created this page. The Kolibri announcement names no authors.

Models & Products

• Kolibri-1 — Apache-2.0 open-weight MoE, 78.1B total / 3.46B active, English and German, released 2026-10-03.

No earlier Aleph Alpha model has a page on this wiki, and none was established as current by anything read. The lab is known to have released models before this one; this page does not state which, because nothing read named them.

Recent Activity

• 2026-10-03: Released Kolibri-1 — 78.1B total parameters, 3.46B active, 384 experts (6 active), 50 layers, Apache 2.0, weights on Hugging Face. Pre-trained on 20 trillion tokens across three stages, of which 4.3 trillion German (21.3% of the mix), ~62% English and ~14% code. Claimed to "match models with up to four times its active parameter count" (source)

Strategic Position

A bid for the sparsity–sovereignty corner, not for the leaderboard. The two things Aleph Alpha chose to publish about Kolibri are an unusually low active parameter count (3.46B of 78.1B) and an Apache-2.0 licence with full weights, and the comparison table runs against two ~120B open models rather than against any frontier system (source). That is a claim about cost-to-run and about who may run it, which is a different axis from the one Frontier Pacing tracks.

The German-language commitment is the part that is measurable rather than rhetorical: 4.3 trillion German tokens, 21.3% of a 20-trillion-token mix, is a far higher non-English share than any other model on this wiki discloses. Set against this, the lab publishes no price, no hosted endpoint and no frontier comparison, so its route to users is stated as download the weights and nothing else.

This section is written from one release. It should be revisited when a second Aleph Alpha artefact is captured here.

Related

• Open-Weights Policy Fight — the licence and the active-parameter count are what place this release there
• Mistral AI — the other European lab on this wiki whose public argument is sovereignty plus open weights
• Test-Time Compute (Inference-Time Compute Scaling) — sparsity and recurrence as efficiency axes