Aleph Alpha's sovereign open-weight Kolibri scores 96.9% on AIME, beating 3x bigger MoE models
TejasKumar_ · x · 2026-10-03
Aleph Alpha released Kolibri, a sovereign open-weight German model (78B total params). Each layer has 384 tiny specialists with a router activating 6 per word, so it effectively runs 3.5B params per token — and scores 96.9% on AIME, beating every MoE model tested including ones 3x bigger; only a dense model doing 8x the work wins. It's self-hostable. Its tokenizer needs 15% fewer tokens than GPT-5's on the German constitution ("bundesverfassungsgericht": 6 tokens vs 2), making German cheaper, faster, and more context-efficient; in their evals it tops every open model its size in both English and German.
Related event: Aleph Alpha Open-Sources Sovereign European Model Kolibri(2 posts)→
More from Models
- No slowdown: Gemini, Bel, Fable and SSI models all launching this month, says account — iruletheworldmo · 2026-10-03
- ChatGPT Plus users report weekly usage cap cut to ~half, from $80-100 to $50 API-equivalent — pensuke89 · 2026-10-03
- Rumor: OpenAI to ship major model next week, originally planned for DevDay — GPT-6.1 Astra? — kimmonismus · 2026-10-03
- Alibaba's Justin Lin: huge sparse models take up memory but won't fix reasoning — JustinLin610 · 2026-10-03
- Vibe coder builds full game in ~24 Opus 5.5 prompts, rates ChatGPT's Astra 2/10 by comparison — NotFromMilkyWay · 2026-10-03
- Rumor: Gemini 4.0 Flash could drop this Thursday at Google's Gemini at Work event — Resident-Pattern731 · 2026-10-03