Aleph Alpha releases Kolibri, a 78B-parameter AI model

by@testingcatalog

Summary

Aleph Alpha has unveiled Kolibri, a notable 78 billion parameter model with 3.46 billion active parameters, and a context window of up to 1 million tokens, distributed under the Apache 2.0 license. This model, which supports reasoning and tool calling, stands out as one of the first general-purpose models from the EU to achieve such performance levels. The development of Kolibri involved a focus on integrating German data, with 21.3% of pre-training tokens being in German, reflecting the broader trend among European AI developers to align with regulations like the EU AI Act and GDPR. The release of Kolibri under an open-weight model allows users to run it on their own hardware, further enhancing accessibility and adaptability for AI applications.

Analysis

Mistral: Mistral is a French AI company known for building competitive open-weight general-purpose models. It is cited in the news as one of the primary European examples of advanced AI development prior to Aleph Alpha's latest release. Aleph Alpha: Aleph Alpha is a German AI lab developing advanced language models with a focus on European data and regulatory compliance. The company recently released its Kolibri model as a general-purpose open-weight offering, highlighting bilingual German-English capabilities and explicit support for reasoning and tool use. This release positions Aleph Alpha as a leading EU contender in high-performance AI alongside other regional labs. Regulation: European AI developers are increasingly designing models from the outset to align with the EU AI Act, GDPR, and related codes of practice. EU Ecosystem: Additional European labs are entering the space with models optimized for regional languages and long-context performance. Open Weights: Releases of high-performance models under the Apache 2.0 license allow users to run and adapt them locally on their own hardware.

Categories

aitechmachine_learning
View Original Tweet