Aleph Alpha releases Kolibri, a 78B English-German model with open weights

TL;DR

• On 3 October Aleph Alpha released Kolibri, an English-German mixture-of-experts model, with full weights on Hugging Face under Apache 2.0.
• It has 78.1B parameters, uses 3.46B per token and serves up to 1,048,576 tokens of context through Aleph Alpha's vLLM plugin.
• Aleph Alpha ran every benchmark itself, and the weights need about 78 GB of GPU memory.