https://localfirstai.eu/tags/exp_037/Recent content in Exp_037 on Local First AIHugoen-usFri, 09 Oct 2026 00:00:00 +0000https://localfirstai.eu/posts/2026-10-09-we-stopped-the-kolibri-benchmark/Fri, 09 Oct 2026 00:00:00 +0000https://localfirstai.eu/posts/2026-10-09-we-stopped-the-kolibri-benchmark/On Monday our own gate refused to certify our MLX port of Aleph Alpha's Kolibri. We rebuilt the gate around forced expert routing and froze it before any run. The port passed in run 2 of three permitted, after a fix to a tokenizer-check cache in our kit keyed by object address, and again in run 3, after a workaround in our runner for a buffer leak in mlx-lm 0.32.0 that Kolibri's layers expose; neither fix touched the port. Then our frozen plan rule, fed with Kolibri's real answer lengths, put even the smallest scored run at 56 hours against a 31-hour laptop budget, and we stopped. There are no Kolibri quality scores here; there are four speed, memory, tokenizer and 4-bit numbers beside their pre-registered thresholds, with no verdicts.