{"slug": "german-ai-startup-aleph-alpha-launches-kolibri-billed-as-a-sovereign-european", "title": "German AI Startup Aleph Alpha Launches Kolibri, Billed As A Sovereign European Model", "summary": "Aleph Alpha released Kolibri, a 78.1-billion-parameter mixture-of-experts English-German language model, on Hugging Face under the Apache 2.0 license on Saturday, positioning it as a sovereign alternative for governments and regulated industries. The German company trained Kolibri from scratch on roughly 24 trillion tokens, with German making up more than 20% of the data mix, using 768 Nvidia B200 GPUs in Germany and Finland; only about 3.46 billion parameters (roughly 4.4%) are active per token. Aleph Alpha says Kolibri leads the MoE models it compared with an overall suite score of 75.5 in English and tops AIME 2025 (EN) at 96.9, as the company is in the middle of being absorbed by Canada's Cohere.", "body_md": "Even as the frontiner of AI moves ever-forward, there’s also a growing number of “sovereign” AI models.\n\nGermany-based Aleph Alpha has released Kolibri, an English-German language model it is positioning as a sovereign alternative for governments and regulated industries. The company put the weights on Hugging Face on Saturday under the Apache 2.0 license, so organisations can download it and run it on their own hardware. In its announcement on X, Aleph Alpha described Kolibri as a small bird with fast wings, a nod to the model’s emphasis on efficiency.\n\nThe launch arrives at an unusual moment for the company, which is in the middle of being absorbed by Canada’s Cohere.\n\n## What Kolibri is\n\nKolibri is a mixture-of-experts transformer with 78.1 billion total parameters, of which only about 3.46 billion (roughly 4.4%) are active for any given token. Each of its 50 layers holds 384 routed experts, with six selected per token alongside one shared expert. Attention is a hybrid: four sliding-window layers for every full-attention layer, which keeps long contexts affordable.\n\nOn context, the model was trained natively up to 262,144 tokens, and Aleph Alpha says it has validated quality and serving efficiency up to 1,048,576 tokens. The company recommends staying at or below 262,144 tokens for latency-sensitive or complex work. Kolibri also has an explicit reasoning mode with four effort settings (none, low, medium and high) and supports tool calling.\n\nAleph Alpha says it trained the model from scratch on roughly 24 trillion tokens across pre-training, mid-training and long-context extension. German makes up more than 20% of the data mix, and the company added over 2 trillion German tokens that it either curated from the web or generated synthetically. Synthetic data is fast becoming standard practice in the industry; Cohere’s CEO Aidan Gomez has said that [the overwhelming majority of the data Cohere generates for new models is synthetic](https://officechai.com/ai/when-human-data-is-too-expensive-synthetic-data-could-be-be-used-to-train-ai-models-cohere-ceo-aidan-gomez/). A custom tokenizer built around German word structure is meant to make German text cheaper to process without hurting English.\n\nPost-training combined supervised fine-tuning on curated German and English data with reinforcement learning across more than 1.2 million internally curated tasks covering reasoning, tool use, instruction following, code and retrieval. The company also trained the model to abstain when the supplied evidence doesn’t support an answer, using what it calls the Merlin-Arthur protocol, a technique aimed at reducing hallucinations in retrieval-based setups.\n\nTraining ran on 768 Nvidia B200 GPUs, on infrastructure in Germany and Finland. That detail matters to the sovereignty pitch: Aleph Alpha says it controlled data curation, training, evaluation and deployment end to end, and it is a signatory of the EU’s general-purpose AI Code of Practice.\n\n## How it performs\n\nAleph Alpha says Kolibri sits on the Pareto frontier of quality versus serving cost in both English and German among the open models it evaluated, and that it can match models with several times as many active parameters. On math and science benchmarks, its comparisons against [Qwen3.6-35B-A3B](https://officechai.com/ai/qwen3-6-35b-a3b-benchmarks/), [Nvidia’s Nemotron 3 Super](https://officechai.com/ai/cheapest-ai-models/) and Mistral Small 4 look like this:\n\n| Benchmark | Kolibri | Qwen3.6 35B-A3B | Nemotron 3 Super | Mistral Small 4 | \n|---|---|---|---|---|\n| AIME 2025 (EN) | **96.9** | 84.6 | 91.7 | 79.8 | \n| AIME 2025 (DE) | 87.5 | 82.9 | 85.6 | 72.3 | \n| AIME 2026 (EN) | **96.0** | 91.0 | 90.4 | 83.1 | \n| AIME 2026 (DE) | **90.0** | 84.4 | 87.5 | 78.5 | \n| GPQA Diamond (EN) | **84.3** | 83.4 | 78.0 | 74.7 | \n| GPQA Diamond (DE) | **81.3** | 80.6 | 76.6 | 72.9 | \n\nAcross its full evaluation suite, Kolibri posts the highest overall score among the MoE models it compared, at 75.5 in English and 70.8 in German. Against its own predecessor, Kolibri Origin, the company reports about 2.7x higher throughput and a gain of more than 21 points in English for the post-trained model, and a 1.6x throughput gain and roughly 23 points for the base model.\n\nThe results aren’t uniformly strong, though, and they are all self-reported. Kolibri trails Qwen on agentic coding, scoring 27.7 on TerminalBench 2.1 against 39.7 for Nemotron 3 Super, and 66.4 on SWE-Bench Verified versus 73.8 for Qwen3.6-35B-A3B. It is also weaker at answering from memory and at multi-turn tool use. Alibaba’s dense Qwen3.8 27B scores higher overall, though it activates several times more parameters per token. The model also needs Aleph Alpha’s own plugin for the vLLM inference server rather than running out of the box.\n\nBecause all 78 billion parameters have to sit in memory even though few are active, the FP8 weights take up about 78 GB. That means a single H200 or B200, or two H100s, is enough to serve it. The open-weights landscape has been [dominated by Chinese models](https://officechai.com/miscellaneous/these-are-the-most-popular-ai-models-on-openrouter-june-2026/) in terms of real-world usage, which is the field Kolibri’s backers hope a European entrant can crack in regulated sectors.\n\nA note on the license: the Apache 2.0 grant covers the published weights and configuration files only. Aleph Alpha retains rights to its training code, architecture and methods.\n\n## About Aleph Alpha\n\nAleph Alpha was founded in Heidelberg in 2019 by Jonas Andrulis and Samuel Weinbach. Andrulis is a serial entrepreneur who founded two AI companies before this one and spent several years at Apple, including in its Special Projects Group and on AI research for Siri. The company built the Luminous family of language models, emphasised transparency and explainability, and sold only to enterprises and governments rather than consumers.\n\nIts funding came in stages. Aleph Alpha raised roughly €5.3 million in early seed money, followed by a €23 million Series A led by Earlybird in 2021. In November 2023 it announced a Series B of more than $500 million from a consortium that included Schwarz Group, the owner of Lidl, along with Bosch Ventures, SAP, Hewlett Packard Enterprise and others. The headline figure drew scrutiny afterwards, since a good portion of it reflected research funding and order commitments rather than fresh equity. The company is regularly listed among [Europe’s most valuable AI startups](https://officechai.com/ai/most-valuable-ai-startups-in-europe-2026/), though it has long sat in the shadow of France’s Mistral, which recently [raised a $3.48 billion Series D at a $24 billion valuation](https://officechai.com/ai/mistral-raises-3-48-billion-series-d-doubles-valuation-to-24-billion/).\n\nOver time, Aleph Alpha stepped back from the race to build frontier models and refocused on sovereign AI deployments, including its PhariaAI platform. Andrulis gave up the CEO role in 2025 and has since left the company; he founded a new startup, CNTR, in February 2026. Ilhan Scheer and Reto Spörri, a former Lidl e-commerce executive, took over management, and the company cut around 50 jobs earlier this year. It now employs about 200 people in Germany.\n\n## The Cohere deal\n\nIn April, Cohere announced it would take over Aleph Alpha, and on September 16 the two signed a definitive merger agreement. The combined company will operate as Cohere and is reportedly valued at around $20 billion. It will be dual-headquartered in Toronto and Berlin, with Heidelberg kept as a research center. Scheer will become Cohere’s chief operating officer and Weinbach its chief research officer. Schwarz Group is committing around $600 million to Cohere’s upcoming Series E, and Cohere plans to offer its technology through Schwarz’s STACKIT cloud. The deal still needs regulatory approval and is expected to close later this year.\n\nGomez has framed the tie-up as a bet on AI that is, in his words, “powerful enough to compete, but secure and governable enough to trust”. Cohere, which has [pitched its research agents to enterprises](https://officechai.com/ai/cohere-ceo-aidan-gomez-explains-how-ai-can-do-a-weeks-worth-of-an-analysts-work-in-30-seconds/) as a way to compress weeks of analyst work into seconds, reported $240 million in annual recurring revenue this year. Capital remains the structural constraint for European players; Eric Schmidt recently argued that [Europe simply can’t raise money at the scale frontier AI demands](https://officechai.com/ai/the-real-limit-to-ai-not-energy-but-cash-former-google-ceo-eric-schmidt/), which is part of why sovereignty-focused specialisation, rather than a head-on frontier race, has become the continent’s more realistic play.\n\nFor now, Kolibri is Aleph Alpha’s statement of what that specialisation looks like. Weinbach said the company tailored the model to German-language use cases so that it can reason in German, a capability aimed squarely at public administration and industry customers. How the model fits into Cohere’s broader lineup once the merger closes is the open question.", "url": "https://wpnews.pro/news/german-ai-startup-aleph-alpha-launches-kolibri-billed-as-a-sovereign-european", "canonical_source": "https://officechai.com/ai/kolibri-aleph-alpha/", "published_at": "2026-10-04 09:04:15+00:00", "updated_at": "2026-10-04 18:41:07.333287+00:00", "lang": "en", "topics": ["large-language-models", "ai-research", "generative-ai", "ai-policy", "ai-infrastructure"], "entities": ["Aleph Alpha", "Kolibri", "Cohere", "Hugging Face", "Nvidia", "Qwen3.6-35B-A3B", "Nemotron 3 Super", "Mistral Small 4"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/german-ai-startup-aleph-alpha-launches-kolibri-billed-as-a-sovereign-european", "markdown": "https://wpnews.pro/news/german-ai-startup-aleph-alpha-launches-kolibri-billed-as-a-sovereign-european.md", "text": "https://wpnews.pro/news/german-ai-startup-aleph-alpha-launches-kolibri-billed-as-a-sovereign-european.txt", "jsonld": "https://wpnews.pro/news/german-ai-startup-aleph-alpha-launches-kolibri-billed-as-a-sovereign-european.jsonld"}}