What Is Kolibri 1?
Kolibri 1 is Aleph Alpha's first open-weight large language model, released on October 3, 2026. It is a bilingual model for German and English — not a multilingual generalist. The model card states the design philosophy directly: depth over breadth, supporting two languages excellently rather than many languages adequately. The weights are published on Hugging Face under Apache 2.0, so anyone can download and use them commercially, while the training code and architecture remain Aleph Alpha's property.
The Architecture: 78B Parameters, 3.46B Active
- Total parameters: 78.1B; active per token: only ~3.46B (about 4.4%)
- 50-layer mixture-of-experts transformer with 384 specialist experts per layer — the router sends each token to 6 routed experts plus 1 shared expert
- Sliding-window attention: 4 of every 5 layers look at just the nearest 512 tokens, keeping long inputs affordable
- FP8 precision weights (half the size of 16-bit) with BF16 for sensitive parts
- 1M-token context length (Aleph Alpha recommends 262K or less for serving efficiency)
- Reasoning mode and tool calling built in
The Honest Benchmark Picture
Aleph Alpha published its numbers unusually frankly. Kolibri 1 scores 84.3 on GPQA Diamond in English (81.3 in German), 96.9 on AIME 2025, 66.4 on SWE-Bench Verified, and 63.2% on RULER at a 1-million-token context. The card's own "best dense model" baseline outscores Kolibri in every category — the company never names that baseline, but it frames Kolibri's argument as efficiency per token rather than raw capability. On its German benchmark suite it averages 71% while decoding roughly 47,000 bytes per second per GPU — nearly double the throughput of the cited baseline at 24,000.
| Specification | Kolibri 1 (FP8) | Kolibri 1 (BF16) |
|---|---|---|
| Total parameters | 78.1B | 78.1B |
| Active per token | ~3.46B | ~3.46B |
| Memory footprint | ~78 GB | ~156 GB |
| Minimum hardware | 2x A100 80GB or 1x H200 | 4x A100 80GB or 1x B200 |
| Knowledge cutoff | June 18, 2026 (EN+DE) | June 18, 2026 (EN+DE) |
Sovereignty Angle: Why Germany Built It
Aleph Alpha positions Kolibri as "sovereign AI made in Germany" — a model you can run entirely on your own infrastructure with no dependency on a foreign provider's servers, built under European and German law with the EU AI Act and GDPR in mind. The company is a signatory of the EU GPAI Code of Practice, and the model card explicitly lists autonomous operation without human review as outside intended use. About 23% of the 20-trillion-token pre-training corpus is German, and the tokenizer (UniBPE) was built around German morphology so German text consumes fewer tokens — a concrete, measurable efficiency gain for German-language workloads.
Prompting Kolibri 1 Effectively
- Use the recommended sampling settings: temperature 1.0, top_p 0.97, top_k 128
- Write German system prompts in German — the model was explicitly optimized for it, and German morphology in the tokenizer reduces token waste
- Keep contexts at or under ~262K tokens for the best efficiency; the 1M window is real but expensive
- Use it for what it is good at: multi-step reasoning, retrieval-augmented generation, agentic tool calling, coding, and German/English assistants
- Skip it for homelab-scale single-GPU setups, non-German/English content, or autonomous coding agents that need maximum raw capability
System: Du bist ein präziser Assistent für deutsches Verwaltungsdeutsch. Antworte auf Deutsch, strukturiert mit Überschriften und Stichpunkten. User: Fasse die folgende E-Mail in drei Sätzen zusammen und extrahiere alle Fristen: [email text]. Reason step by step in English internally, answer in German.
Frequently Asked Questions
Is Kolibri 1 free to use commercially?
The model weights are published under Apache 2.0, which permits commercial use, including self-hosted deployment. But it is open-weight, not fully open source — Aleph Alpha keeps the rights to training code, architecture, and training methods, and you supply your own GPU hardware or rental costs.
What hardware do I need to run Kolibri 1?
The FP8 variant needs about 78 GB of GPU memory (minimum 2x A100 80GB, 2x H100, 1x H200, or 1x B200). The BF16 variant needs about 156 GB (minimum 4x A100 80GB, 4x H100, 2x H200, or 1x B200). This is data-center or rented-cloud hardware, not a home PC.
How is Kolibri 1 different from Mistral's models?
Kolibri is bilingual (German + English) by design rather than multilingual, with a German-optimized tokenizer and 23% German in its training data. Its pitch is sovereign, self-hosted operation under EU law with efficiency-first MoE architecture — roughly 3.46B active parameters per token out of 78.1B total.
Can Kolibri 1 generate images or handle voice?
No. It is a text-only model with no image, audio, or video input or output. For multimodal work you need a different model; Kolibri covers text generation, reasoning, coding, and agentic tool calling.
Is Kolibri 1 better than the biggest proprietary models?
Not on raw capability — the model card's own unnamed dense baseline beats it on every benchmark category. Its advantage is cost-per-token efficiency, German-language depth, and full self-hosting under Apache 2.0, which makes it competitive for agentic and coding workloads rather than frontier research tasks.
