{
  "id": 11700095,
  "title": "Kolibri Has Landed: A Sovereign Open-Weight Model",
  "url": "https://urgent.news/2026/10/03/kolibri-has-landed-a-sovereign-open-weight-model",
  "topic": "ai",
  "section": "AI",
  "published": "2026-10-03T09:36:04.000Z",
  "source": {
    "name": "Hacker News",
    "slug": "hacker-news",
    "url": "https://aleph-alpha.com/en/blog/kolibri-has-landed-a-sovereign-open-weight-model/"
  },
  "original_language": "en",
  "account": "On the day of German reunification, Kolibri has landed - a sovereign open-weight model. This English-German mixture-of-experts Transformer boasts 78B total parameters, with 3B active. It supports context lengths up to 1 million tokens and can be downloaded with full weights from Hugging Face under the Apache 2.0 license. Kolibri evolved through continuous model training efforts, beginning with the Kolibri Origin model, a 30B total, 3B active model with a 65k token context window. Like Kolibri Origin, Kolibri underwent the same pipeline: data ingestion and curation, ablations, pre-training, post-training, and final evaluations. This pipeline enabled running numerous ablation experiments and stable pre-training without needing human intervention for hardware failures or data connection drops. Kolibri's development prioritized optimization for sovereign mission-critical work in regulated areas such as public administration, industrials, and aerospace. Specialization for German, reasoning, math, agentic behavior, and other customer-specific needs in production resulted in contextualized performance improvements, enabling customers to monitor the economic impact and measure growing return on investment. Kolibri's sovereignty combines two dimensions: the model's construction and its transferability to customers. Full supply-chain integrity is provided, with transparency for customers' freedom of deployment and intellectual property safety. Optimized for performance across various sectors, Kolibri's small and efficient size allows for on-premise deployment without sharing internal data with third-party inference services. The model sits on the Pareto frontier for quality versus serving cost, delivering top quality at lower costs compared to other models. Kolibri's internal evaluation suites, developed specifically for sectors like the German public sector, aviation, manufacturing, and automotive industry, outperform public benchmarks. To ensure accuracy in answering questions, Kolibri was trained with abstention data and the Merlin-Arthur protocol, enabling it to say \"I don't know\" when the answer isn't in the context. Native German and English model with 21.3% pre-training tokens in German, developed using sparing translation to maintain cultural authenticity. Developed in Germany, trained on infrastructure in Germany and Finland, under EU AI Act, General-Purpose AI Code of Practice, and GDPR guidelines, Kolibri's entire pipeline is under control, ensuring full sovereignty. By owning the end-to-end pipeline, customers enjoy flexibility in deployment and controllable reasoning effort to balance cost, latency, and answer quality.",
  "summary": null,
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 2,
    "also_reported_by": [
      {
        "outlet": "Hacker News Best",
        "title": "Kolibri Has Landed: A Sovereign Open-Weight Model",
        "url": "https://urgent.news/2026/10/03/kolibri-has-landed-a-sovereign-open-weight-model-11736792",
        "published": "2026-10-03T09:36:04.000Z"
      }
    ]
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}