What "open" covers here

Kolibri's weights and configuration files are downloadable from Hugging Face under Apache 2.0. Download them, fine-tune them, build a commercial product on top of them. Aleph Alpha keeps its training code and data pipeline private, which is its right, and worth knowing: open weights are not reproducible science. Still, measured against the recent wave of "open" releases that quietly exclude whole regions from the license, like MiniMax did with H3, Apache 2.0 is the real thing.

The hardware clause

This is the part the announcement states softly. Kolibri is a mixture-of-experts model: 384 experts per layer, six routed experts plus one shared expert awake for each token. Compute scales with the 3.46 billion active parameters. Memory does not. The full 78 billion parameters have to sit somewhere, and in FP8 that is about 78 gigabytes. Aleph Alpha's own listed minimum is two 80 GB A100s, or one H200, B200, or B300. Community builds have squeezed it onto big-memory Macs and CPU-only PCs with enough RAM, but "open" and "runs on your laptop" are two different claims. Only the first is on offer.

The million-token asterisk

The release advertises context up to 1,048,576 tokens. The longest training stage ran at 262,144; the million figure is validated rather than native, with Aleph Alpha providing serving settings to reach it. The release benchmarks are, as always, the vendor's own. Long context that degrades quietly past the trained range is a familiar failure mode. Treat the million-token claim as a testable promise, not a proven fact, until someone independent runs it end to end.

Sovereignty as the actual product

Kolibri was trained on infrastructure in Germany and Finland, under German and EU law, with about a fifth of its pretraining tokens in German. The sales pitch is twofold: your data never leaves your hardware, and the compliance paperwork already speaks EU AI Act. For a German public agency, that combination is the product. The model is the delivery vehicle. Announced on German Unity Day, which tells you exactly which audience it is aimed at.

Sources

  1. [1] TestingCatalog — “Aleph Alpha releases open-weight Kolibri with 1M context”Read source
  2. [2] Denny Sentinel — “The Model Is Open, the Memory Bill Is Not”Read source
  3. [3] AlexTech AI — “Aleph Alpha opens a German AI model with a million-token context”Read source