Aleph Alpha Releases Kolibri: A 78.1B Open-Weight English-German MoE Model With Only 3.46B Active Parameters

Aleph Alpha has unveiled Kolibri, a significant advancement in open-source language models designed to balance capability with computational efficiency. The 78.1-billion-parameter Mixture-of-Experts (MoE) model represents a strategic approach to deploying advanced AI on limited hardware resources. With an Apache 2.0 license, Kolibri's open weights make it freely accessible for both research and commercial applications, democratizing access to sophisticated multilingual AI technology.

Kolibri employs a Mixture-of-Experts architecture that activates only 3.46 billion parameters per token, despite its massive 78.1-billion-parameter foundation. This sparsity-based approach significantly reduces computational demands during inference while maintaining model expressiveness. The model supports a 1-million-token context window, enabling processing of extensive documents and complex conversations. Kolibri handles both English and German, addressing a substantial multilingual AI gap in open-source models. Critically, the model's FP8-quantized weights run efficiently on single-GPU setups, specifically NVIDIA's B200 or H200 accelerators, eliminating the need for expensive multi-GPU infrastructure.

• Accessibility: Open-weight release under Apache 2.0 removes licensing barriers for enterprises and researchers worldwide
• Cost Efficiency: Reduced active parameters dramatically lower inference costs and enable deployment on consumer-grade hardware
• Multilingual Support: English-German capabilities address specific market needs while demonstrating scalable multilingual design
• Context Length: The 1-million-token window enables advanced applications in document analysis, long-form content generation, and extended conversations
• Per-Request Reasoning: Configurable reasoning effort allows users to optimize speed versus quality trade-offs dynamically
• Hardware Democratization: Single-GPU deployment removes infrastructure barriers that previously limited model access

The release of Kolibri marks a crucial shift toward efficient, democratized AI infrastructure. As organizations increasingly demand cost-effective AI deployment, models that achieve competitive performance with minimal computational overhead become essential. By combining open weights, multilingual capabilities, extended context, and hardware efficiency, Kolibri positions itself as a practical alternative to proprietary models for businesses seeking autonomy and cost control. This development signals growing momentum in the open-source AI community to deliver enterprise-grade models without proprietary constraints, potentially reshaping how organizations approach language model deployment and development.