Aleph Alpha releases open model Kolibri for German government agencies: it runs on the customer's own servers
Prepared by Sintelligent with the help of SI. Publication time restored from the build log. . Updated: .
European governments are getting their own open SI that they can run at home; for now it is noticeably weaker than the leaders, but the choice is widening.
Kolibri is a 78.1-billion-parameter model, of which 3.46 billion are active per token (a mixture-of-experts architecture), with a context of up to 1 million tokens; it was trained on 768 Nvidia B200 GPUs, and the weights are open on Hugging Face under the Apache 2.0 license. The company targets it at public administration, industry and aerospace, where it matters who sees the data, and claims the model is comparable to ones with four times as many active parameters. According to Trending Topics' estimate, on the Artificial Analysis index it would likely score 15–20 points and trail a dozen or more open models from Alibaba, Z.ai, Moonshot and Xiaomi; Aleph Alpha replies that public benchmarks poorly reflect customers' needs.