Kolibri 1 TL5
Kolibri 1 is Aleph Alpha's German-developed, open-weight text reasoning model for industrial and public-sector applications. Its 78B-parameter Mixture-of-Experts architecture activates 3.46B parameters per token, pairing efficient inference with a context window validated up to 1M tokens (256K native and recommended for complex tasks). Trained from scratch on a bilingual corpus with about 24% German text, it uses a German-optimized tokenizer for efficient processing of compound words. It supports controllable reasoning, native tool calling, coding, structured extraction, and retrieval-augmented generation (RAG), including document-grounded answers and multi-step agents. Training-data sourcing and governance are documented in a public technical report. Developed and trained in Europe, Kolibri 1 is released as open weights under Apache 2.0.
Kolibri 1 TL5
Kolibri 1 is Aleph Alpha's German-developed, open-weight text reasoning model for industrial and public-sector applications. Its 78B-parameter Mixture-of-Experts architecture activates 3.46B parameters per token, pairing efficient inference with a context window validated up to 1M tokens (256K native and recommended for complex tasks). Trained from scratch on a bilingual corpus with about 24% German text, it uses a German-optimized tokenizer for efficient processing of compound words. It supports controllable reasoning, native tool calling, coding, structured extraction, and retrieval-augmented generation (RAG), including document-grounded answers and multi-step agents. Training-data sourcing and governance are documented in a public technical report. Developed and trained in Europe, Kolibri 1 is released as open weights under Apache 2.0.
