What this is

This week, German company Aleph Alpha open-sourced Kolibri-1 on Hugging Face: 78 billion total parameters, but only 3.46B active per token (the smallest unit of text a model processes) — a textbook signature of MoE (Mixture of Experts) architecture: big model, cheap to run.

The context window reaches 1 million tokens (roughly the length of War and Peace), and the license is Apache 2.0 (allowing commercial use, modification, and re-closing the source — one of the most enterprise-friendly open-source licenses).

Aleph Alpha has long positioned itself as Europe's "sovereign AI" play (domestic AI solutions that don't depend on the U.S. tech stack), with backing from German retail giant Schwarz Group — the parent company of Lidl.

Industry view

We noticed engineers on Reddit's LocalLLaMA community and Hugging Face circles are split on this one.

The positive take: 3.46B active parameters means inference cost (the actual compute consumed during a Q&A) sits near a mid-sized small model, while the knowledge capacity comes from 78B. This is the first open-source model to combine "big parameters + small active" with million-scale context — a signal for enterprises eyeing on-premise deployment.

But the pushback is sharp:

  • 78B MoE VRAM (GPU temporary storage; memory the model occupies at runtime) requirements aren't low — running 1M context takes a stack of H100/A100-class GPUs, far from "run it on your laptop"
  • Aleph Alpha's valuation has been marked down since its last funding round, commercialization has stalled over the past year, and the open-source release looks more like "had to do it" than "strategic choice"
  • No complete third-party benchmarks yet — whether the parameters are well-stacked will be decided by leaderboard scores, not PDF reports

Impact on regular people

For enterprise IT: If your company is evaluating private deployment (installing AI on your own servers), MoE + million-token context is becoming the new "price-performance reference line." When procuring compute, "total parameters" is no longer the core metric — "active parameters × context length" is.

For individual careers: MoE won't directly change your daily work, but it will change the price of AI tools. Once open-source MoE models mature, enterprise-tier ChatGPT-class subscriptions (currently several thousand to tens of thousands per user per year) will face downward pressure.

For the consumer market: Consumers won't touch Aleph Alpha's models anytime soon. But the company targets European governments and large enterprises — if this playbook works, it means Chinese large-model companies competing for B-end (enterprise customers) and G-end (government customers) deals face not just American options, but a European one too.