Aleph Alpha has released Kolibri, an open-weight mixture-of-experts language model featuring 78.1B total parameters and 3.46B active parameters. Designed for sovereign deployment in regulated sectors, the model supports a 1M-token context window and is available under an Apache 2.0 license.
The release of Kolibri provides a high-efficiency alternative for organizations requiring strict data sovereignty under the EU AI Act and GDPR. By utilizing a sparse architecture that activates only 3.46B parameters, Aleph Alpha enables long-context processing that is computationally cheaper than dense competitors like Qwen or Mistral. This efficiency allows streaming and media firms to process vast archives of bilingual metadata or legal documentation on localized hardware without relying on US-based cloud providers. As the industry moves toward agentic workflows, Kolibri’s high scores in English and German RAG benchmarks suggest a viable path for specialized European enterprise applications. Watch for adoption rates within German public administration as a signal for sovereign AI viability.
The model's focus on EU transparency compliance aligns with broader regional efforts to secure AI-generated media pipelines. This development complements recent enterprise AI agent tuning initiatives aimed at optimizing model performance in regulated environments.
Aleph Alpha has launched Kolibri, a bilingual English-German mixture-of-experts model featuring 78.1B parameters. By activating only 3.46B parameters per token, it offers a cost-effective, high-efficiency solution for long-context processing. This release is significant for organizations needing to maintain strict data sovereignty under the EU AI Act and GDPR regulations.
The Kolibri model supports a massive 1-million-token context window, specifically 1,048,576 tokens, enabled by a hybrid attention mechanism.
Yes, the Kolibri model is available under an Apache 2.0 license.
Kolibri uses a sparse mixture-of-experts architecture that activates only 3.46B of its 78.1B parameters per token, making it computationally cheaper than dense models while supporting long-context processing.
Deployment is optimized for single-GPU environments, specifically running on NVIDIA B200, B300, or H200 hardware via vLLM.
Add StreamingMeme as a preferred source on Google to see more of our streaming news at the top of your Search results.
Add as preferred source