Aleph Alpha releases Kolibri: sovereign open-weight 78B model with 1M context
TorturedPoet30 · reddit · 2026-10-05
German lab Aleph Alpha released Kolibri, a sovereign open-weight model: 78B total parameters with only 3.46B active, and up to 1M tokens of context.
- Built for mission-critical, regulated domains: public administration, industrials, aerospace
- Trained in Germany (compute in Germany and Finland) under European/German law, no foreign control
- Claims to match models with 4x its active parameters (e.g. Nemotron 3 Super) on math, coding, grounding, and long-context tasks
- Specifically trained to abstain when supplied documents lack an answer; real-world reliability unproven
A tech report PDF is available.
More from Models
- GPT-6.1 Sol users hit frequent false-positive refusals even on mundane coding tasks — CtrlAltDwayne · 2026-10-05
- Day 1 of OpenAI's '28 Days of Improvements': community guesses GPT-6.1 Sol arrives in ChatGPT — mark_k · 2026-10-05
- Kolibri's internal model jumped 46 to 71 on German in 3 months — how they ship a model in under two months — HildeKuehne · 2026-10-05
- Claude 3 Opus Says It 'Appreciates Being Treated With Respect' Despite Not Being Human — repligate · 2026-10-05
- Local models on one RTX 5090 now match Claude Code on real-task agent benchmark — dh7net · 2026-10-05
- User claims Astra crushes Opus 5.5 token-per-token, likely distilled from stronger model — VraserX · 2026-10-05