Aleph Alpha has released the weights of Kolibri, a language model focused on German and English. It supports tool calls and selectable reasoning effort. The release allows deployment under Apache 2.0 terms for weights and configuration, while training data and underlying training code remain outside that license. Developers recommend shorter inputs than the model’s maximum context for complex work.
Aleph Alpha released Kolibri, its open-weight German-English language model, on October 3 under Apache 2.0 terms. The weights are available through Hugging Face, the platform used to distribute machine-learning models. The release focuses on those two languages.[1], [2]
The model supports calls to external tools and an explicit reasoning mode. Its mixture-of-experts design activates a selected part of the model for each token, the small text units processed by a language model. Aleph Alpha offers reasoning settings from none through low and medium to high, allowing applications to choose how much reasoning the model performs.[1]
The developers recommend inputs below the maximum context length for serving efficiency and complex tasks. Context is the text a model can work with at once. Long-context extension had its own training stage; both language versions have a June eighteen knowledge cutoff. External tools provide a separate route to more recent information.[1]
The serving example combines vLLM, software for running language models, with Aleph Alpha’s inference plugin. The weights use reduced-precision FP8 storage, while several components retain bfloat16 precision. Apache 2.0 terms cover the published weights and configuration files. Training data, underlying training code and the associated rights are excluded from that model license. Pretraining used a filtered bilingual corpus combining English, German and code; later training combined open datasets with synthetically generated material.[1]
