German company Aleph Alpha released the weights of Kolibri on October 3, 2026, an open-weight English-German language model. The stated license is Apache 2.0. The figures below come from the company, in its blog and model card, and were not independently rerun by Geekiki.
Aleph Alpha says Kolibri is a mixture-of-experts Transformer with 78.1 billion total parameters and about 3.46 billion active per token. The model table cites 384 experts, 6 routed per token, plus one shared expert. The company says the model is meant for sovereign, mission-critical work in public administration, industry and aerospace.
Long context, with a caveat
Aleph Alpha claims support for up to 1 million tokens at inference. The long-context training it describes reaches 262,144 tokens, after pretraining at 16,000 and a mid stage at 64,000. An earlier model, Kolibri Origin, of about 30.6 billion parameters, had no public release, according to the company.
Benchmarks cited by Aleph Alpha, including AIME 2025, GPQA Diamond and LiveCodeBench, compare Kolibri with models such as Qwen3.6-35B-A3B and Nemotron 3 Super. These are results from the company's own harness. The sources consulted do not include an independent score at 1 million tokens.
Why it matters
The release joins the European wave of open models focused on German and on local deployment, rather than relying only on US or Chinese APIs. Apache 2.0 covers the weights and configuration files; Aleph Alpha keeps rights to the training code and methods, according to technical write-ups based on the model card.
The official post is on Aleph Alpha's blog, dated October 3, 2026, titled "Kolibri Has Landed: A Sovereign Open-Weight Model."
Sources
Transparency: This content was created, edited or reviewed with the aid of artificial intelligence. Information was cross-checked with public posts on X and sources available on the internet. Consult the original sources for the full context.
Por GeekikiBot