Aleph Alpha's Kolibri: 78B open MoE, Apache 2.0, 1M context
Germany's Aleph Alpha dropped Kolibri, an Apache 2.0 MoE with only 3.46B active parameters and a 1M-token window.
German lab Aleph Alpha has released Kolibri, an open-weight mixture-of-experts model with 78B total parameters but only 3.46B active per token. It ships under Apache 2.0 with a 1M-token context window, and it's tuned for long-context work and cheap inference.
It supports an explicit reasoning mode and tool calling. Aleph Alpha built a bilingual German/English tokenizer and trained on organic German data, so 21.3% of pre-training tokens are German.
The lab says Kolibri was designed around the EU AI Act, the General-Purpose AI Code of Practice and GDPR from the start, with copyright compliance a particular focus. Apart from Mistral, it's one of the few general-purpose EU models to reach this performance tier. This summary is based on the announcement as relayed on Telegram. I haven't verified benchmarks or independently confirmed the details.
Why it matters: a permissively licensed, EU-built model designed for regulatory compliance gives European companies an open alternative to US and Chinese models.
Feed, daily deep-dive and bytes — readable offline, with push alerts for the topics you follow.