What happened
Aleph Alpha released Kolibri, a German-English model with 78 billion parameters, about three billion active per token, trained on 768 B200 GPUs in Germany and Finland and available under Apache 2.0 on Hugging Face.
Why it matters
Aleph Alpha claims Kolibri sits on the Pareto front of quality and operating cost in both languages, outperforming compared models with similar architectures on either metric.
What to watch
The Pareto front claim depends on how Kolibri performs against comparable models in practice; watch whether independent evaluations confirm its 71% score on German benchmarks and its decoding speed versus GPT OSS A5B, Qwen 3.6 A3B, and Gemma 4 A4B.
WHO IT HITSPublic administration, aviation, and industry teams evaluating open-weight models for German-language use are the target adopters, since Aleph Alpha says Kolibri was developed under European law with the EU AI Act in mind.
Summaries like this, in your inbox every morning.
Aleph Alpha built a dedicated German data pipeline for this project, and German accounts for 21.3 percent of the training data. Chinese models were also used to generate synthetic training data. The model was developed under European law with the EU AI Act in mind. These choices reflect a deliberate approach to training data and regulatory context that the company connects to the model's design.
The reported benchmark figures and speed comparisons give a concrete sense of where Kolibri stands relative to other models. Whether those results hold up under broader testing, and whether the target sectors adopt an open-weight model on this basis, are the open questions the release leaves on the table.
Pick your industry and the AI tools you use, and get news related to your work every day.
Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.
Ghost AI raised $11 million, led by Andreessen Horowitz with Abstract, Audacious Ventures, Nova and SV Angel…
OpenAI will roll out invisible text watermarks to all ChatGPT and Codex plan users in the EU within weeks, and…
Reflection announced Beam, a 501B-parameter open-weight model with 23B active parameters, claiming parity with…
A Stanford, Carnegie Mellon, UC Berkeley, and Microsoft Research team ran 6,800+ math, coding, and science tas…
Reflection AI launched Beam, a 501 billion-parameter open-source LLM
A step-by-step guide fine-tunes Muse Glimmer, Meta's 30B vision model, locally for equation-to-LaTeX conversio…
