https://www.opensourceway.blog/tags/kolibri-model-card-no-score-published-on-own-bias-benchmark/Recent content in Kolibri-Model-Card-No-Score-Published-on-Own-Bias-Benchmark on 「开源之道」Hugozh-CNCopyright (c) 2016 - 2026, 「开源之道」·适兕; all rights reserved.Wed, 28 Oct 2026 01:33:12 +0800https://www.opensourceway.blog/en/posts/china-open-source-daily/2026/10/2026-10-28/Wed, 28 Oct 2026 01:33:12 +0800https://www.opensourceway.blog/en/posts/china-open-source-daily/2026/10/2026-10-28/On 2026-10-03, the Day of German Unity, Aleph Alpha released Kolibri — 78.1B parameters, Apache 2.0, and, on its own account, trained from scratch in Germany and Finland with no base-model dependencies. Then page 53 of the 189-page technical report names the models that wrote the reasoning traces: GLM-5.2 and GLM-5.3 from Z.ai, and Qwen3.8-27B from Alibaba. Section 3.1.1.5 concedes the point in the company's own words — 'we acknowledge that several of our teacher models were built in China, so they carry known political biases on sensitive topics.' Twenty-five days earlier the same company had published a study measuring exactly that: six Chinese models returned a balanced answer to 17–41 percent of 967 politically sensitive prompts. Verified for this briefing: the study publishes no score for Kolibri itself, the model card carries a warning that Kolibri 'may reproduce political biases present in its training data,' and a third-party reading has found that the study tested no model from Z.ai — the lab behind two of the three teachers. The open-weight regime's defining property is not that weights are free. It is that the chain of custody is written down — because a licence cannot be negotiated in a way that only works if a ministry asks first.