Kolibri: Aleph Alpha Ships a 3B-Active Open Model That Says I Don't Know
Germany's flagship AI lab just shipped the thing it was always supposed to ship. Aleph Alpha released Kolibri on Saturday, an English-German Mixture-of-Experts model with 78 billion total parameters and only 3 billion active per token. Apache 2.0, weights on Hugging Face. It hit the top of Hacker News twice in one day, 441 points for the launch post and 409 for an outside teardown, plus a third thread on the tech report.
The numbers punch far above the active size. AIME 2025 at 96.9%, AIME 2026 at 96.0%, HumanEval+ at 92.7%. Aleph Alpha says it matches models with up to 4x its active parameters on math, coding and long context. Context goes to 1M tokens. The architecture is 384 experts with 6 active, 50 layers, 40 of them sliding-window attention and 10 full attention, which is how you keep a million tokens cheap. Training took 21 days on 768 B200s across about 24 trillion tokens.
The most agent-relevant number is the boring one. On AA-Omniscience it abstains 44% of the time, against 14.8% for its predecessor. The model was trained to say "I don't know" when the answer isn't in the context. For an agent that will take actions based on what it believes, a model that refuses to bluff is worth more than two points on a math benchmark. Tool calling and four reasoning effort levels (none to high) ship out of the box.
"Sovereign" is the sales pitch to European governments and regulated industries, and 21.3% German pre-training tokens is the proof. But 3B active means it runs on modest hardware, and that's what will decide whether developers outside Germany pick it up for local agents.
Link: huggingface.co/Aleph-Alpha/Kolibri-1 and aleph-alpha.com/en/blog/kolibri-has-landed-a-sovereign-open-weight-model
← Back to all articles
The numbers punch far above the active size. AIME 2025 at 96.9%, AIME 2026 at 96.0%, HumanEval+ at 92.7%. Aleph Alpha says it matches models with up to 4x its active parameters on math, coding and long context. Context goes to 1M tokens. The architecture is 384 experts with 6 active, 50 layers, 40 of them sliding-window attention and 10 full attention, which is how you keep a million tokens cheap. Training took 21 days on 768 B200s across about 24 trillion tokens.
The most agent-relevant number is the boring one. On AA-Omniscience it abstains 44% of the time, against 14.8% for its predecessor. The model was trained to say "I don't know" when the answer isn't in the context. For an agent that will take actions based on what it believes, a model that refuses to bluff is worth more than two points on a math benchmark. Tool calling and four reasoning effort levels (none to high) ship out of the box.
"Sovereign" is the sales pitch to European governments and regulated industries, and 21.3% German pre-training tokens is the proof. But 3B active means it runs on modest hardware, and that's what will decide whether developers outside Germany pick it up for local agents.
Link: huggingface.co/Aleph-Alpha/Kolibri-1 and aleph-alpha.com/en/blog/kolibri-has-landed-a-sovereign-open-weight-model
Comments