# Aleph Alpha Releases Kolibri: Apache 2.0 German/English MoE, 3.46B Active

Times of AI Desk · 2026-10-03 · Models

[https://timesof.ai/2026/10/aleph-alpha-kolibri-open-weight](https://timesof.ai/2026/10/aleph-alpha-kolibri-open-weight)

> Aleph Alpha’s October 3 release, Kolibri, is an English-German mixture-of-experts model with 78.1 billion total parameters and 3.46 billion active, context up to about 1 million tokens (the model card recommends serving at or below 262,144), and full weights on Hugging Face under Apache 2.0. The company says it was trained in Germany and Finland and that 21.3% of pre-training tokens are German. Scores such as AIME 2025 English 96.9 are Aleph Alpha’s own harness at high reasoning effort, not an Arena or Artificial Analysis placement.

A European lab just put a bilingual open-weight model on Hugging Face under **Apache 2.0**. The capability claim is still the lab’s own scoreboard.

**Aleph Alpha**, in an **October 3, 2026** post dated to the Day of German Reunification, says **Kolibri** is an English-German **mixture-of-experts** transformer with **78 billion total parameters** and about **3 billion active**, **context up to 1 million tokens**, downloadable in full on Hugging Face under **Apache 2.0**. The **Hugging Face card for `Aleph-Alpha/Kolibri-1`** is more exact: **78,103,074,560** total parameters, **3,457,573,120** active per token (**3.46 billion**), release date **October 3, 2026**, license **Apache 2.0**, knowledge cutoff **June 18, 2026** for both English and German. The card says quality was validated up to **1,048,576** tokens and **recommends at most 262,144** for serving efficiency and complex tasks. Weights are **FP8**, with a stated memory footprint of **about 78 GB**.

## What actually shipped

| | Kolibri, as published |
| --- | --- |
| Repo | `Aleph-Alpha/Kolibri-1` |
| Total / active | 78.1B / 3.46B (card); blog prose says ~3B active |
| Context | Up to 1,048,576; card recommends ≤262,144 |
| License | Apache 2.0 on the weights |
| Reasoning | none, low, medium, high; tool calling |
| Knowledge cutoff | EN and DE, June 18, 2026 |
| German data | Blog: **21.3%** of pre-training tokens |
| Where trained | Blog: Germany and Finland, under European law |
| Pre-training compute | Card: **768 NVIDIA B200s**, **21 days** (pre-training only) |

The blog’s build table matches the card on the 3.46B active figure, four reasoning efforts, and a **native long-context train at 262,144** tokens, with the million-token length treated as an extension. Serving is not stock vLLM alone: both the post and the card require Aleph Alpha’s **`aleph-alpha-inference`** plugin (`vllm serve Aleph-Alpha/Kolibri-1` with Kolibri reasoning and tool parsers).

Aleph Alpha’s sovereignty pitch is deployment control plus a German/English data mix, not a new legal status. The blog says the model is aimed at public administration, industrials, and aerospace, that customers can run it on-premise, and that **21.3% of pre-training tokens are German** (about **4.3 trillion** of a **20 trillion** token run). The card’s training-data summary describes the filtered bilingual corpus with a **German share nearer 24%**; that is a mix description, not a second measured benchmark. This piece uses **21.3%** for the share of tokens the blog says the run actually saw.

## Claims vs checks

Aleph Alpha says benchmarks used **its own harnesses** and, where applicable, **each model’s highest reasoning effort**. On that table, Kolibri’s **AIME 2025** is **96.9 English** and **87.5 German**; **GPQA Diamond** English is **84.3**; **LiveCodeBench v6** is **85.9**. The same table’s **AA-Omniscience Index (public set)** for Kolibri is **−32.8** — a lab number, not a placement on Artificial Analysis. The blog’s line that Kolibri “matches models with up to four times its active parameter count, such as Nemotron 3 Super” is **that comparison**, not an independent audit.

The card’s eval note says category averages are unweighted means and that Kolibri was run at **reasoning effort high**. Those settings are not a license to rank Kolibri against closed frontier models on Arena or Artificial Analysis.

A public-board check on the **evening of October 4, 2026**, did **not** show Kolibri in the visible **Artificial Analysis** Intelligence Index table or on the **LM Arena** page opened then. That is a snapshot of what was on screen, not a search of every alias, and it is not a quality verdict.

## Limits

- All benchmark figures above are **Aleph Alpha harness results**. They were not re-run here.
- Blog prose says “~3B active”; the card and the blog’s own spec table say **3.46B**. The exact card figure is the one used for size.
- German-share wording differs between the blog’s **21.3% of pre-training tokens** and the card’s corpus-mix description. They are not averaged.
- “Pareto frontier” for quality versus serving cost is **the company’s framing**.
- Arena / Artificial Analysis: **not observed** on the October 4 evening snapshot. That check was not refreshed later the same evening. No Elo is inferred.
- Weights were not downloaded; license, size, and release date are what the **blog and model card** state.

## Sources

- [Aleph Alpha: Kolibri Has Landed: A Sovereign Open-Weight Model (October 3, 2026)](https://aleph-alpha.com/en/blog/kolibri-has-landed-a-sovereign-open-weight-model/)
- [Hugging Face: Aleph-Alpha/Kolibri-1 model card (release date October 3, 2026)](https://huggingface.co/Aleph-Alpha/Kolibri-1)
- [Artificial Analysis: models leaderboard (evening snapshot October 4, 2026; Kolibri not in the visible Intelligence Index table)](https://artificialanalysis.ai/leaderboards/models)
- [LM Arena leaderboard (evening snapshot October 4, 2026; Kolibri not on the opened page)](https://lmarena.ai/leaderboard)