Expansion of llama.cpp Support to Include Aleph Alpha’s Kolibri-1
The open-source machine learning ecosystem has seen a technical expansion with the addition of support for Aleph Alpha’s Kolibri-1 model within the llama.cpp framework. According to a recent pull request submitted to the ggml-org/llama.cpp repository, the implementation aims to enable local execution of the European-developed model.
Technical Specifications of Kolibri-1
Kolibri-1 is a Mixture-of-Experts (MoE) reasoning model developed by Aleph Alpha Research GmbH. The architecture utilizes a 50-layer transformer design with 384 experts per layer, consisting of one shared expert and six routed experts. While the model possesses a total of 78.1 billion parameters, its MoE structure ensures that only 3.46 billion parameters are active per token, optimizing inference efficiency.
The model is designed for bilingual performance in German and English. According to technical documentation from Aleph Alpha, the model was trained on a 20-trillion-token filtered corpus, comprising approximately 62.5% English and 23.9% German data. The training process included a mid-training phase of 5 days and a long-context extension phase of 13 hours, utilizing significant computing resources, including 768 NVIDIA B200 nodes.
Capabilities and Hardware Requirements
Kolibri-1 features an explicit reasoning mode and supports tool calling, making it suitable for multi-step reasoning, coding, and retrieval-augmented generation (RAG). It supports a context length of up to 1,048,576 tokens, though the developer recommends a limit of 262,144 tokens for optimal serving efficiency.
The hardware requirements for running the model in its FP8 precision format are substantial. The model memory footprint is approximately 78 GB. Aleph Alpha recommends high-end hardware configurations for deployment, such as two NVIDIA H100 SXM5, two H200, or a single B200 or B300 unit. Minimum requirements include two NVIDIA A100 80 GB or two H100 SXM5 units.
Implementation Status
The integration into llama.cpp was reported via a pull request on Reddit, where a contributor noted that while support for the model architecture has been added, the community is currently awaiting the release of GGUF-formatted files to facilitate immediate use.
As of the report, the model is released under the Apache 2.0 license. Aleph Alpha has also noted that the company is a signatory of the EU GPAI Code of Practice. The model’s knowledge cutoff for both English and German datasets is June 18, 2026.