Aleph Alpha releases Kolibri 1, a 78B German and English reasoning MoE under Apache 2.0
Kolibri 1 is a 78B mixture-of-experts reasoning model with 3.46B active parameters, tuned for German and English, under Apache 2.0.
Aleph Alpha published Kolibri 1, a mixture-of-experts reasoning model focused on German and English. The card gives 78.1B total parameters, 3.46B active per token, 50 layers, and 384 experts per layer with one shared and six routed. Context is 1,048,576 tokens, with 262,144 recommended for serving. It supports an explicit reasoning mode and tool calling.
The card gives a release date of 3 October 2026 and a knowledge cutoff of 18 June 2026. Pre-training used 20T tokens, about 62.5 percent English, 23.9 percent German and 13.6 percent code, on 768 NVIDIA B200 GPUs over 21 days. Weights are FP8, around 78 GB, and the card lists a minimum of two A100 80 GB cards or one H200.
The weights are published under Apache 2.0. A technical report and a tech blog are linked.
A European model with its hardware bill and training log printed on the card. That is the useful kind of sovereign.
Sources
Editor's first pass (AI-assisted)
- Real 4: Apache 2.0, ungated FP8 weights; the card lists minimum hardware such as 2x A100 80 GB.
- Useful 3.5: German and English reasoning, RAG and tool calling with an explicit reasoning mode.
- Notable 3.5: 962 likes; a European developer's sovereign-model pitch.
- Novel 3: Bilingual focus and a Muon-trained 384-expert MoE; the architecture is otherwise familiar.
Comments