Aleph Alpha Open-Sources Kolibri: 78B Parameters, Only 3.46B Active per Token

On October 3, German AI company Aleph Alpha released Kolibri — deliberately chosen to coincide with German Unity Day. Kolibri is an open-weight model under the Apache 2.0 license, with weights published on Hugging Face, free to download and use commercially. The name “Kolibri” is German for “hummingbird,” paired with the tagline “Small bird, fast wings, Kolibri is here.”

Specs: 78B Total, 3.46B Active

The architecture is MoE (mixture of experts). Total parameters: 78.1B, but only 3.46B activate per token — about 4.4%. The structure: 50 layers, each with 384 experts plus one shared expert, with the router picking 6 experts per token. Compute cost scales with 3.5B parameters; VRAM footprint scales with 78.1B — roughly 78GB in FP8.

Context: 260K Native, Expandable to 1M

The context window reaches 1,048,576 tokens — about a million. Native training length is 262,144, and the company provides configuration to extend it to the full million. Long documents, lengthy reports, entire codebases in a single pass.

Training: 768 B200s, 24 Trillion Tokens

Training ran on 768 NVIDIA B200s, with compute in Germany and Finland. Total training data: roughly 24 trillion tokens — 20T pre-training, 3.44T mid-training, 200B long-context adaptation. German makes up 21.3% of pre-training data; the 128K-token English-German vocabulary was built specifically for German compound words. Knowledge cutoff: June 18, 2026. The company also published a 189-page technical report covering architecture, data, and evaluations in detail.

On performance, the official numbers: 96.9% on AIME 2025, 96.0% on AIME 2026, 84.3% on GPQA Diamond. The comparison models in its class are Qwen3.6-35B-A3B and Mistral Small 4. Discussion on Hacker News flagged a wrinkle: under Kolibri’s own eval harness, Qwen3 27B actually scored better in German — and wasn’t included in the official comparison.

Inference intensity comes in four tiers (none, low, medium, high), adjustable per request, with tool calling supported — suited for agentic workloads.

Positioning: Sovereign AI for Local Deployment

The target customers are governments, public administration, industry, aerospace, and defense — places where data is sensitive and compliance demands are high. The logic: institutions run it in their own data centers, so sensitive documents never go to a third-party API. Training ran entirely in Germany and Finland, designed around the EU AI Act and GDPR, with the data pipeline anonymized first. Minimum hardware: one B200 / B300 / H200, or two H100 SXM5s, or two A100 80GBs, with vLLM and the official inference pack. Not for ordinary PCs — this is built for institutions.

Kolibri didn’t come from nowhere. Aleph Alpha first built a smaller Kolibri Origin (~30B total parameters, 3B active, 65K context) to validate the pipeline, then spent about three months from June to September scaling up to the full release.

European labs rarely release open weights at this scale. What sets Kolibri apart isn’t the parameter count — it’s the posture: training, compliance, and deployment engineered end-to-end for European regulation. In the open-weight race, there’s now a player selling “compliance” as the feature.