Aleph Alpha released Kolibri, an open-weight bilingual model. It supports English and German with a 1M token context window. The model uses 78.1B parameters with 3.46B active per token. Weights are available under Apache 2.0 on Hugging Face. It can be deployed on-premises for regulated industries. Kolibri offers four reasoning modes and tool calling. Training used 24T tokens across Germany and Finland. It targets sovereign, mission-critical government and industry use. Benchmarks show strong math, code, and grounding performance. Aleph Alpha claims it matches models with four times the active parameters. Grounding uses Merlin-Arthur to withhold answers without evidence. Deployment requires Aleph Alpha’s inference package and vLLM plugin.

• 78.1B parameters, 3.46B active per token
• 1M token context window
• 24T tokens total training data
• 75.5 English score, 70.8 German score