Aleph Alpha's Kolibri: The Open (Apache 2.0) Model You Run on Your Own Server | inicIA
Direct answer
On October 3, 2026, Germany's Aleph Alpha released Kolibri 1: an open-weight language model under the Apache 2.0 license (download, modification, and commercial use permitted). The numbers: 78.1 billion total parameters, with only 3.46 billion active per token (mixture-of-experts architecture: the model activates a few "experts" per token instead of the whole network), optimized for English and German, with 262,144 tokens of native context (validated up to 1,048,576). It was trained on infrastructure in Germany and Finland, under European and German law. It is the most serious bet yet on sovereign AI: running the model on your own infrastructure instead of sending your documents to an outside provider.
What to know before getting excited
- The weights are free; running it is not: the FP8 model is about 78 GB, and the listed minimum is data-center hardware — 2× 80GB NVIDIA A100s, 2× H100s, or 1× H200/B200/B300. This is not software you install on your laptop.
- The context, honestly: 262,144 tokens is the native context recommended for efficient serving; the million-token figure is validated, but Aleph Alpha recommends staying at or below 262K for complex tasks.
- It reasons and uses tools: four reasoning effort levels (none, low, medium, high) plus tool calling in the standard format.
- It admits what it does not know: Aleph Alpha says it trained the model with a proprietary method (Merlin-Arthur protocol) to recognize when an answer is not in the documents it was given, instead of inventing one.
Who it makes sense for
- Regulated companies (finance, healthcare, government, legal): where sending customer data to an external API is a compliance problem, not a preference.
- Operations where sensitive data is the asset: if your competitive advantage lives in your documents, keeping the model inside your perimeter removes a whole risk surface.
- Scale with predictable cost: at high, sustained volume, fixed infrastructure cost can beat per-token API billing — run the math on your real numbers.
The sovereign-AI checklist
- Define what data never leaves: classify your information before choosing architecture. What cannot leave determines whether you need on-premise.
- Calculate total cost: hardware + energy + staff to run it vs. per-token API at your volume. Free weights are not free AI.
- Demand sector evaluations: Aleph Alpha publishes evaluation suites for public administration, automotive, semiconductors, and industry. Ask the same of any provider.
- An update plan: a self-hosted model goes stale. Define who evaluates it and when it gets replaced.
Related links
- Google Opened Gemini 4 Argon With Launch Pricing: $2 per Million Tokens Until Nov 30
- Privacy and Data: What Happens to Your Customers' Information
- The War of Cheap Verdicts: The AI Layer That Decides "Yes or No" Without Writing an Essay
Frequently asked questions
What is Aleph Alpha's Kolibri?
An open-weight language model released Oct 3, 2026: 78.1B total parameters (3.46B active per token), English and German, 262K tokens of native context, Apache 2.0 license permitting commercial use.
Can I use it free in my business?
The weights are free and the license permits commercial use, but you need serious infrastructure to run it (data-center GPUs, ~78GB in weights alone). For most, the practical path is a provider that operates it for you.
What is sovereign AI?
Running AI models inside your own infrastructure and legal jurisdiction, without depending on foreign clouds or sending your data to third parties. Kolibri was trained in Germany and Finland under European law.
How does it differ from using an API like OpenAI's or Anthropic's?
Full control over data and deployment, fixed vs. variable cost, and simpler compliance in regulated sectors — in exchange for you (or your provider) operating the infrastructure.
Take the next step
If your data cannot leave your company — by regulation or by strategy — sovereign AI is now a real option with open weights. Talk to inicIA on WhatsApp: https://wa.me/17866120930
Chat on WhatsAppBy inicIA
