Aleph Alpha has released Kolibri, a German-English language model with downloadable weights and local deployment options, following training on Verda’s European infrastructure. The company is targeting public administration, industrial and aerospace workloads with the October 3 release.
Kolibri is a mixture-of-experts model with about 78 billion total parameters and 3.46 billion active parameters per token. It supports tool calling and lets users choose among four reasoning settings, from no reasoning to low, medium and high effort. The weights and configuration files are available under the Apache 2.0 license.
The model card lists a maximum context window of 1,048,576 tokens, but recommends using no more than 262,144 for serving efficiency and complex tasks. The card lists a June 18 knowledge cutoff for both German and English; tool use can provide newer information.
Aleph Alpha says Kolibri was built for deployments where customers need control over their infrastructure and data. Its account of sovereignty includes visibility into the model-building process and freedom to deploy the resulting model. Those are the developer’s descriptions of its design and commercial offering, rather than an independent certification of every deployment’s compliance.
In an October 5 account, Verda said training took place in European data centers that it owns and operates. It supplied a bare-metal fleet using Nvidia Blackwell GPUs and InfiniBand networking, and maintained cluster services for Aleph Alpha. The infrastructure provider also built Kubernetes and observability systems and contributed engineering work on training efficiency.
DeepSeek and Huawei have separately expanded software support for Ascend processors, seeking a domestic Chinese computing stack around different hardware.
Aleph Alpha’s technical account describes Kolibri training on 768 B200 GPUs. It says the pipeline automatically handled 38 unplanned interruptions during 21 days of pre-training, restarting jobs on another set of nodes and resuming from checkpoints. The company attributes those interruptions to hardware faults or connection timeouts.
The release materials also describe training intended to make the model abstain when supplied documents lack the evidence needed for an answer. Performance and hallucination-reduction results in those materials are company-run evaluations. The model card continues to warn about incorrect outputs and bias, requiring users to assess suitability for their applications.
The card puts the model’s FP8 weight memory footprint at roughly 78 gigabytes and lists supported configurations including a single H200, B200 or B300 GPU. The Apache license applies to the published weights and configuration files; the card excludes underlying code, architecture and training methods from that license grant.
Sources: Aleph Alpha, Aleph Alpha, Verda
–
By the Control Plane Editorial Team