Aleph Alpha has released Kolibri, a new open-weight large language model designed specifically for German and English language processing. Released on October 3, 2026 under the Apache 2.0 license, Kolibri represents what the company describes as a 'sovereign' AI model built with training infrastructure in Germany and Finland under European and German law. At its core, Kolibri is a mixture-of-experts (MoE) model with 78.1 billion total parameters that activates only approximately 3.5 billion parameters per token processed. This efficiency is achieved through 50 layers, each containing 384 expert sub-networks plus one shared expert, with a routing mechanism that directs each token to six of the 384 experts. According to Aleph Alpha, this approach allows the model to deliver capabilities comparable to much larger models while reducing computational requirements per token. The model introduces several technical innovations aimed at improving German language handling. Its tokenizer, trained with a novel UniBPE algorithm, uses 128,000 tokens and requires 11.2% fewer tokens for German text compared to GPT-5's tokenizer, based on Aleph Alpha's measurements. This efficiency is particularly pronounced with legal German text, where Kolibri needed 15% fewer tokens than GPT-5's tokenizer when processing the Basic Law for the Federal Republic of Germany. Kolibri employs a hybrid attention mechanism where 40 of its 50 layers use sliding-window attention (each token attends to the previous 512 tokens), while every fifth layer uses full attention across all previous tokens. This design enables the model to handle extended contexts affordably, with native support for 262,144 tokens and tested capabilities up to 1,048,576 tokens. To address language-specific reasoning challenges, Aleph Alpha trained Kolibri on approximately 800,000 German reasoning examples, resulting in German math scores of 87.5 on the 2025 American Invitational Mathematics Examination (AIME) - outperforming comparable models with similar active parameter counts. The model also implements what Aleph Alpha calls the 'Merlin-Arthur protocol' to improve uncertainty handling, training it to recognize when it lacks sufficient information to provide an answer rather than hallucinating. Performance evaluations by Aleph Alpha show Kolibri leading open models of its size in several areas: it achieves 75.5 overall score in English and 70.8 in German, exceeds comparable models on AIME 2025 mathematics tests (96.9 in English, 87.5 in German), and demonstrates strong performance on long-context tasks at 1 million tokens (63.2 on the RULER benchmark). The model also shows capability in answering questions over unseen company documents in English (89.7 score). However, Aleph Alpha acknowledges several limitations. Kolibri performs worse on closed-book knowledge tests, scoring 51.0 on the Retrieval-Augmented Generation Benchmark's questions without supporting documents - significantly below the 22.2% correct rate of comparable models. Multi-turn tool calling appears weaker, with a score of 39.8 on the Berkeley Function Calling Leaderboard's multi-turn tests compared to 58.2 for GLM-4.7 Flash. Coding agent performance is also modest, with 27.7 on Terminal-Bench 2.1 and 66.4 on SWE-bench Verified. Practical deployment considerations include substantial memory requirements - approximately 78 GB of weights in 8-bit floating point format, necessitating data-center GPUs such as pairs of NVIDIA A100/H100 cards or single H200/B200/B300 processors. The model requires Aleph Alpha's specific vLLM plugin (currently supporting vLLM version 0.29) for operation, and as of launch, no hosted providers were serving the model. Kolibri's deliberate focus on just two languages (German and English) represents a trade-off where depth in these languages is prioritized over broader multilingual coverage. The model's knowledge cutoff is June 18, 2026, and it supports four levels of reasoning effort (none, low, medium, high) adjustable per request, along with tool calling capabilities. Aleph Alpha has signed the European Union's General-Purpose AI (GPAI) Code of Practice, positioning Kolibri as compliant with upcoming AI regulations.