Aleph Alpha launches Kolibri with 3.46 billion active parameters per token
Kolibri uses a mixture-of-experts architecture with 78.1 billion total parameters, activating 3.46 billion for each token. Its weights are available under the Apache 2.0 license. According to MarkTechPost, the English-German model also offers a context window of one million tokens and lets users set reasoning effort for each request.
- RuntimeWire reports 78 GB of FP8 weights.
- RuntimeWire says the model card calls for at least two A100 80GB GPUs.
- MarkTechPost says the FP8 weights run on a single B200 or H200.
Covered by 2 publishers within 15 hours of the first report.
Reporting2
RuntimeWire Aleph Alpha releases German AI model with 78 GB of FP8 weights · Ryan Merket
MarkTechPost Aleph Alpha Releases Kolibri: A 78.1B Open-Weight English-German MoE Model With Only 3.46B Active Parameters · Asif Razzaq