Germany's sovereign open-weight Kolibri: 78B params, 3.5B active, 96.9% on AIME 2025
TejasKumar_ · x · 2026-10-04
Aleph Alpha released Kolibri, a sovereign open-weight (Apache 2.0) German model: 78B total params with only 3.5B active per token, scoring 96.9% on AIME 2025 — beating every tested MoE model, some with 3x more active params. Key details: 384 experts per layer with a router picking 6; 78GB memory (2x H100 or 1x H200); mostly 512-token attention with full attention every 5th layer enabling 1M-token context; trained on 800k German reasoning examples after finding sparse German data hurt; trained to admit "I don't know". Its tokenizer needs 15% fewer tokens than GPT-5's on the German constitution.
More from Models
- Daniel Han publishes summary of LLM benchmarks you can actually trust — danielhanchen · 2026-10-06
- Claim Verification Benchmarks Mostly Test Retrieval, Not Reasoning, Finds 24K-Trace Study — deliprao · 2026-10-06
- COLM26 study: LLMs ace claim verification benchmarks by taking shortcuts, not verifying — deliprao · 2026-10-06
- Opus 5.5 uses 26k tokens vs Astra's 12k yet costs 23% less per task at equal AA score — ChrisGPT · 2026-10-06
- GPT-6 Astra claimed to be first AI crossing world-class astrophysics threshold — johnseach · 2026-10-06
- $500/mo AI subscription is huge money in Jakarta: PPP pricing debate — sujingshen · 2026-10-06