Aleph Alpha releases Kolibri open-weight German-English model

Aleph Alpha released Kolibri, an Apache 2.0 English-German MoE with about 78.1B parameters and 3.46B active.

Aleph Alpha released Kolibri, also referred to as Kolibri 1, on 3 October 2026, an Apache 2.0 open-weight English-German mixture-of-experts model for on-premise sovereign deployments in German public administration, industrials, and aerospace. Sources disagree slightly on size: one gives 78 billion total parameters and 3 billion active, while the more specific account gives 78.1 billion total and about 3.46 billion active per token. Context is described as up to 1 million tokens in one report and, more precisely, as a native 262,144-token window tested to 1,048,576 in the other. Training used roughly 24 trillion tokens, more than a fifth of them German, on 768 NVIDIA B200 GPUs in Germany and Finland, with an 18 June 2026 knowledge cutoff. Aleph Alpha says the model was built with the EU AI Act in mind and that it signed the EU GPAI Code of Practice; some training data was rephrased with Gemma 4 and Mistral-NeMo, and Qwen3-32B labeled quality filters. Reported scores are AIME 2025 at 96.9, GPQA Diamond at 84.3, and LiveCodeBench v6 at 85.9, and it follows Kolibri Origin, a 30B-total, 3B-active model with a 65k context.

• Aleph Alpha released Kolibri (also called Kolibri 1) on 3 October 2026; full weights are on Hugging Face under Apache 2.0.
• Scale differs slightly by source: 78 billion total and 3 billion active parameters, versus the more specific 78.1 billion total and about 3.46 billion active per token.
• Context is a native 262,144-token window tested to 1,048,576; one report summarizes this as up to 1 million tokens.
• Training used roughly 24 trillion tokens, more than a fifth German, on 768 NVIDIA B200 GPUs in Germany and Finland; knowledge cutoff is 18 June 2026.
• Reported scores: AIME 2025 96.9, GPQA Diamond 84.3, and LiveCodeBench v6 85.9, said to be competitive with models using up to four times as many active parameters.
• It follows Kolibri Origin, a 30B-total, 3B-active model with a 65k context, and targets on-premise sovereign use in German public administration, industrials, and aerospace.
• Aleph Alpha says it was built with the EU AI Act in mind and signed the EU GPAI Code of Practice; some data was rephrased with Gemma 4 and Mistral-NeMo, and Qwen3-32B labeled quality filters.
• A German tokenizer is reported to use 11.2% fewer tokens than GPT-5.