Aleph Alpha releases Kolibri-1: 78B MoE model with 3.46B active parameters for German and English
{'Kolibri-1 is Aleph Alpha’s open-weight mixture-of-experts model with 78.1 billion parameters, designed for German and English. It activates only 3.46 billion parameters per token, enabling efficient inference while supporting up to 1,048,576 tokens of context. The model was trained on 24 trillion tokens, including over 2 trillion German-specific data, and adheres to EU AI Act and GDPR…
Key points
- Aleph Alpha’s Kolibri-1 is a 78.1B MoE model activating only 3.46B parameters per token for German/English tasks.
- Supports 1M-token context with sliding-window attention and UniBPE tokenizer optimized for German compound words.
- Trained on 24T tokens (2T+ German-specific), compliant with EU AI Act/GDPR, and distributed under Apache 2.0 license.
The story so far
5 episodes →- Aleph Alpha releases Kolibri-1: 78B MoE model with 3.46B active parameters for German and Englishthis story
Aleph-Alpha/Kolibri-1 · Hugging Face - 78B parameters. 3.46B active. Up to 1M tokens of context
huggingface.co · 3 October 2026
Loading the full article…
This text was published by huggingface.co. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
Coverage and discussion
5sources- Hacker News discussion · 122 pointsnews.ycombinator.com
- Reddit discussionreddit.com
- Aleph Alpha releases Kolibri, an open-weight model that makes the case for European AI sovereigntyPress · The Decoder ·
- OpenAI Safety Official Resigns, Self-Hosted AI Options EmergePress · note.com ·
- Aleph Alpha Releases Kolibri: A 78.1B Open-Weight English-German MoE Model With Only 3.46B Active ParametersPress · MarkTechPost ·
- Show HN: Germany's new sovereign AI model KolibriPress · tej.as ·
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Generative AI & Models
All →- OpenAI faces concerns over loop transformers; Microsoft confirms GPT-6.1 Sol uses them · 1 src
- Meta introduces Muse Wonder and Muse Wonder Full models · 8 src
- Anthropic releases Claude Haiku 5.5 at lower prices than GPT-6 Luna · 1 src
- Microsoft releases Microsoft-Decision-1, a Qwen3.5-9B-based scoring model · 2 src
- Mistral releases Mistral Large 4, nicknamed Le Chonk, a 1-trillion-parameter multimodal model · 1 src
Comments
via GitHub Discussions