Kolibri's internal model jumped 46 to 71 on German in 3 months — how they ship a model in under two months
HildeKuehne · x · 2026-10-05
The Kolibri team (@IlhanScheer) reveals Kolibri Origin, the internal predecessor to the released Kolibri 1.0: with the same active compute per token, its German score went from 46 to 71 in three months.
The team credits the gains not to luck but to "building the machine that builds the model":
- Automate everything you can
- Measure constantly
- Change your mind fast and stay disciplined
They've documented the full process of shipping a model in under two months. Kolibri 1.1 is also teased as arriving "faster than you think." A worthwhile look at high-iteration training methodology from a small team.
More from Models
- Google doesn't need the best model: a faster, cheaper Gemini Pro stays competitive — haider1 · 2026-10-05
- Claude asks user whether to pull an all-nighter or finish the task tomorrow — prasenx · 2026-10-05
- OpenAI's Decisions API announced at DevDay still has no pricing or docs a week later — Balance- · 2026-10-05
- 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
- Claude 3 Opus Says It 'Appreciates Being Treated With Respect' Despite Not Being Human — repligate · 2026-10-05