Germany Advances Digital Sovereignty with Aleph Alpha's Sovereign AI Model
Berlin, Sunday, 4 October 2026.
Aleph Alpha released Kolibri, a 78.1-billion-parameter open-weight model engineered for regulated European sectors, featuring an anti-hallucination protocol that actively refrains from answering when source evidence is lacking.
Technical Architecture and Specifications
The Kolibri model utilizes a Mixture-of-Experts (MoE) architecture featuring 78.1 billion total parameters, with only 3.46 billion active parameters per token [1][2]. This configuration means that approximately 4.43 percent of the model’s parameters are active during inference, optimizing computational efficiency while maintaining scale [3][6]. The system supports a native context window of 262,144 tokens, which is expandable to 1,048,576 tokens via specific configuration settings [3][6]. To handle this extensive context, the architecture employs 50 layers, with sliding-window attention used in 40 layers and full attention applied every fifth layer [6][8].
Training Infrastructure and Data Composition
Training was conducted using 768 NVIDIA B200 GPUs located in Germany and Finland, ensuring compliance with European data sovereignty laws [2][8]. The model was trained on a corpus of approximately 24 trillion tokens, comprising 20 trillion for pre-training, 3.44 trillion for mid-training, and 200 billion for long-context adaptation [2][3]. The data mix consists of roughly 62% English, 21.3% German, and 14% code, with specific augmentation for German language reasoning [1][3]. A custom tokenizer named UniBPE was developed, featuring a vocabulary of 128,000 tokens to optimize morphological processing for German compound words [2][8].
Sovereignty and Regulatory Compliance
Kolibri is designed for on-premises deployment to align with the European AI Act, GDPR, and the General-Purpose AI Code of Practice [1][2]. The model weights are released under the Apache 2.0 license, though usage is restricted regarding terrorism, violence, and criminal activities under EU Regulation 2024/1689 [3][6]. By training on infrastructure in Germany and Finland under European law, Aleph Alpha ensures no foreign control over the supply chain or data [2][8]. This approach allows government agencies to avoid third-party inference services, maintaining full intellectual-property safety [6][8].
Performance Benchmarks and Capabilities
In evaluations, Kolibri achieved an overall score of 75.5 in English and 70.8 in German, with specific industry RAG performance reaching 99.0% accuracy in automotive supplier contexts [1][3]. On the AIME 2025 math benchmark, the model scored 96.9%, and 84.3% on GPQA Diamond, narrowly beating competitors like Qwen3.5 35B-A3B in German overall scores [4][6]. However, coding performance remains mixed, with a 66.4% score on SWE-Bench Verified, trailing behind Qwen3.6’s 73.8% [4][8]. The model also features four reasoning effort levels, allowing users to adjust thinking depth from none to high [2][3].
Hallucination Reduction Protocols
To address reliability, Aleph Alpha implemented the Merlin-Arthur protocol, which trains the model to refrain from answering when context is insufficient for verification [2][3]. This game-theoretic approach involves hiding parts of the context during training to encourage the model to admit ignorance rather than hallucinate [3][8]. On the Artificial Analysis Omniscience test, Kolibri admitted ignorance or provided partial answers 44% of the time, compared to 11.1% for Qwen3.5 35B-A3B [8]. This feature is critical for public sector applications where accuracy and trustworthiness are paramount [1][6].
Deployment Requirements and Availability
Deployment requires the aleph-alpha-inference package version 1.0 or higher and a compatible vLLM plugin for serving [3][8]. Hardware requirements include a minimum memory footprint of approximately 78 GB, necessitating configurations such as two NVIDIA A100 80 GB GPUs or a single H200 [3][6]. The model is available for download on Hugging Face, with recommended sampling parameters including a temperature of 1.0 and top_p of 0.97 [3][8]. As of October 4, 2026, hosted provider support remains limited, requiring users to manage infrastructure independently [6][8].
Market Implications and Future Developments
The release signifies a major step for European digital sovereignty, offering an alternative to proprietary foreign models for regulated industries [1][5]. Aleph Alpha is currently contemplating scaling up beyond the current 78 billion parameter architecture, though no specific deadline has been provided [2][6]. Industry observers note that the ability to train high-performing models is no longer confined to US or Chinese entities, indicating a shift in the global AI landscape [4][5]. Ongoing projects include utilizing Kolibri for AI-driven neurological triage systems in Berlin, highlighting its potential in specialized vertical applications [8].