Aleph Alpha released Kolibri weights, an Apache-2.0 license, a vLLM plugin and a training report.
Aleph Alpha opens Kolibri for German-English reasoning
2026-10-03Frontier watchTRIAL
Aleph Alpha released Kolibri under Apache-2.0, a German-English reasoning MoE with 78B total and 3.46B active parameters. It supports reasoning levels, tool calling and up to 1M tokens, while recommending at most 262K for efficient serving and complex tasks. Weights, a vLLM plugin and a detailed training report are public; evaluations are author-reported.
What this means for you
For German-language enterprise systems, Kolibri provides a rare inspectable model tuned for German morphology, tool use and long documents. Public serving code makes a bilingual RAG and tool-calling comparison practical before considering adoption.
FRONTIER WATCH
Public artifacts and validated results are different.
The recorded fact is the public work and its artifacts. Author-reported outcomes do not establish independent reproduction. Next checks include usable code, task conditions and third-party validation.
Evidence & related material
Technical details & scores
| Architecture | 78B MoE, 3.46B active; 384 experts per layer with 1 shared and 6 routed |
|---|---|
| Languages | German and English |
| Context | trained to 262,144; validated to 1,048,576; vendor recommends <=262,144 for efficiency |
| Capabilities | reasoning effort, tool calling, structured extraction, RAG |
| License | Apache-2.0 |
| Serving | aleph-alpha-inference vLLM plugin; approximately 78 GB FP8 weights |
| Training | 20T pre-training tokens plus 3.44T mid-training and 201B long-context extension |
Technical impact 3
Engineering value 4
Adoption signal 3
Maturity 2
Verification cost 4
Risk 2