https://localfirstai.eu/tags/exp_038/ Recent content in Exp_038 on Local First AI https://localfirstai.eu/images/og-default.png https://localfirstai.eu/images/og-default.png Hugo en-us Sat, 10 Oct 2026 00:00:00 +0000 https://localfirstai.eu/posts/2026-10-10-kolibri-scored-below-gemma-4/ Sat, 10 Oct 2026 00:00:00 +0000 https://localfirstai.eu/posts/2026-10-10-kolibri-scored-below-gemma-4/ We put Aleph Alpha's Kolibri-1, untrained and at 4-bit, into CasaSol's property-law answering pipeline in place of gemma4:26b, and changed nothing else. On 60 sealed questions it scored 0.10 lower on a 0–2 correctness scale (95 % CI −0.20 to 0.00): level in English, behind in Polish and Spanish. On a new German-buyer slice it showed no edge. Blind judges rated its German native on 13 of 36 answers against gemma's 30, because it carries English words over from the English source text. It fits a 64 GB Mac mini. It does not earn a live-bot trial.