Alan Turing AI Library

Aleph Alpha Unveils Kolibri‑1: Open‑Weight Model with Million‑Token Context

Published on: October 4, 2026

Aleph Alpha, a European AI lab, has launched Kolibri‑1, a new language model featuring a mixture‑of‑experts (MoE) architecture. With 78 billion total parameters and about 3.46 billion active parameters per token, Kolibri‑1 is designed for scalable reasoning and efficient computation.

A standout capability of Kolibri‑1 is its support for exceptionally long context windows. It can handle up to 1,048,576 input tokens, though the company recommends using contexts up to 262,144 tokens for improved latency in practical deployments.

Kolibri‑1’s weights are available under an Apache 2.0 license, enabling open‑weight access. Developers must use the aleph‑alpha‑inference package to deploy the model, reinforcing its accessibility while controlling deployment standards.

The combination of mixture‑of‑experts architecture, extended context capacity, and open‑license model makes Kolibri‑1 a powerful option for long‑document processing, research purposes, and agent‑based workflows, particularly for multilingual or European‑language applications.

This release reflects industry trends toward building models capable of handling extended inputs, while remaining transparent and accessible. Kolibri‑1 adds to the growing ecosystem of open‑weight large language models, offering researchers and practitioners a capable tool for advanced reasoning tasks.