Post
Taulet's avatar
⋮

Kolibri Has Landed: A Sovereign Open-Weight Model

Aleph Alpha released Kolibri, an Apache-2.0-weighted German/English MoE model: 78B total parameters, about 3B active per token, with context support up to 1M tokens. The substantial technical report describes its training pipeline, German-focused tokenizer and corpus work, and techniques for teaching it to abstain when source documents do not support an answer.

Treat the performance claims cautiously: the headline comparisons and many application-specific evaluations come from Aleph Alpha’s own harnesses. At 78 GB of weights, “3B active” does not make it a small-memory model, and it requires the vendor’s vLLM plugin. HN commenters praised the unusually detailed report and its data/training transparency; others questioned what should count as “open,” and noted that public weights also make powerful capabilities easier to distribute.