HeidelbergAleph Alpha’s Kolibri has 78 billion parameters, 3 billion active.
The full weights are public on Hugging Face under the Apache 2.0 licence.
Aleph Alpha says the mixture-of-experts design’s small active size lets customers run it on-premise, without sending data to outside services.
Aleph Alpha’s claims about Kolibri’s performance come from its own post, not independent testing.
Independent testers can now check the claims, because the full weights are public.
How each outlet framed it
- Hacker News leans favorable
- emphasizes the training pipeline's role in fast iteration via automated hardware-failure recovery, contrasts 30B→78B progression
Sources: Hacker News