Aleph Alpha releases Kolibri-1, a 78B English and German model, under Apache 2.0
Only about 3.5 billion of its parameters are active per token, but all 78 billion have to be held in memory.
Aleph Alpha released Kolibri-1 on 3 October 2026, the Day of German Reunification, with the weights on Hugging Face under the Apache 2.0 licence. It is a mixture-of-experts model with 78.1 billion parameters in total and 3.46 billion active for each token, built for English and German, with a tokenizer trained for both and about 21% German in its pre-training data.
Aleph Alpha says it was trained on up to 262,144 tokens of context, and supports up to one million with extra serving settings. The announcement says the model requires the company’s own inference package, a plug-in for vLLM, so it is not a ready-made Ollama model. It does not state file sizes or the hardware needed to run it.
The benchmark figures, including 96.9 on AIME 2025 and 84.3 on GPQA Diamond, come from Aleph Alpha’s own test setups and have not been independently checked. The company says Kolibri matches models with up to four times its active parameters.
Sources
This is a summary of reporting published elsewhere. Where a claim matters, follow the source — it is the record, this is not.
Related on this site
Which models can your PC run?
Scan your hardware and see every model that fits, with the memory each one needs.