{"slug": "aleph-alpha-releases-kolibri-a-78b-parameter-open-weight-ai-model-built-in", "title": "Aleph Alpha releases Kolibri, a 78B-parameter open-weight AI model built in Europe", "summary": "Aleph Alpha released Kolibri on October 3, 2026, a 78.1-billion-parameter open-weight Mixture-of-Experts model with a context window of up to 1,048,576 tokens, trained on 768 NVIDIA B200 GPUs in Germany and Finland. The Heidelberg company tuned Kolibri for English and German, trained it on 20 to 24 trillion tokens, and shipped it under the Apache 2.0 license on Hugging Face, requiring roughly 78 GB of storage in FP8 format. Aleph Alpha reported Kolibri scored 96.9% on AIME 2025 and published a 189-page technical report, targeting regulated sectors including public administration, industry, aerospace, and defense.", "body_md": "# Aleph Alpha releases Kolibri, a 78B-parameter open-weight AI model built in Europe\n\nThe Heidelberg company's Mixture-of-Experts model supports a 1 million token context window and targets regulated sectors in English and German\n\nAleph Alpha has released Kolibri, a 78.1-billion-parameter AI model with open weights and a context window of up to 1,048,576 tokens.\n\nThe Heidelberg-based company built it in Europe, trained it with a heavy emphasis on German content, and pitched it at sectors where compliance matters most: public administration, industry, aerospace, and defense.\n\nThe launch came on October 3, 2026.\n\n## What Kolibri actually is\n\nKolibri is a Mixture-of-Experts transformer, usually shortened to MoE. Kolibri has 78.1 billion total parameters, but only approximately 3.46 billion are active per token.\n\nThe headline feature is the context window. Kolibri is designed to work across up to 1,048,576 tokens, commonly rounded to 1 million.\n\nThe model is tuned for English and German. Its training run covered 20 to 24 trillion tokens, with German content given particular weight.\n\n### AI, tech, and the markets they move—in one daily briefing.\n\nDaily. Free. Join 34,000+ readers across crypto, finance, and policy.\n\n## The hardware, the license, and the fine print\n\nAleph Alpha trained Kolibri on infrastructure located in Germany and Finland, using 768 [NVIDIA](https://cryptobriefing.com/markets/nvidia/) B200 GPUs.\n\nKolibri ships under the Apache 2.0 license and is available on Hugging Face.\n\nKolibri requires roughly 78 GB of storage in FP8 format.\n\nAleph Alpha also published a 189-page technical report covering Kolibri’s architecture, training methods, and evaluation results.\n\nOn performance, Kolibri scored 96.9% on AIME 2025, a math competition benchmark, according to Aleph Alpha’s own reporting.\n\n**Disclosure:** This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our\n\n[Editorial Policy](https://cryptobriefing.com/editorial-policy/).", "url": "https://wpnews.pro/news/aleph-alpha-releases-kolibri-a-78b-parameter-open-weight-ai-model-built-in", "canonical_source": "https://cryptobriefing.com/aleph-alpha-releases-kolibri-ai-model/", "published_at": "2026-10-03 18:39:27+00:00", "updated_at": "2026-10-03 19:07:46.025534+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "generative-ai", "ai-research", "ai-products"], "entities": ["Aleph Alpha", "Kolibri", "NVIDIA", "Hugging Face", "Diego Almada Lopez", "AIME 2025"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/aleph-alpha-releases-kolibri-a-78b-parameter-open-weight-ai-model-built-in", "markdown": "https://wpnews.pro/news/aleph-alpha-releases-kolibri-a-78b-parameter-open-weight-ai-model-built-in.md", "text": "https://wpnews.pro/news/aleph-alpha-releases-kolibri-a-78b-parameter-open-weight-ai-model-built-in.txt", "jsonld": "https://wpnews.pro/news/aleph-alpha-releases-kolibri-a-78b-parameter-open-weight-ai-model-built-in.jsonld"}}