According to VChK-OGPU, Yandex may face issues not only with servers but also regarding the classification of its Alice neural model as "sovereign." A former employee provided screenshots proving that the architecture of Alice’s new flagship "Expert" mode is far less independent than claimed; the assistant likely relies on the Chinese GLM model to answer questions.
Yandex’s official stance is that 80% of Alice AI queries are processed by a sovereign LLM, with requests routed to third-party models only when task complexity increases.
In practice, however, a different pattern emerges. This is evidenced by a test involving the "glitch token" `.DataGridViewColumnHeadersHeightSizeMode`—a known blind spot for the GLM family of models.
This token triggered the same system failure in Alice’s "Expert" mode that occurs in Chinese models: the model fails to recognize the string, cannot repeat it, and asks for clarification on the query.
While GPT-6.1-Sol and Claude Opus 5.5 handle this prompt normally, GLM-5.2 exhibits this specific failure—a reaction mirrored when using another well-known glitch prompt.
Ultimately, Alice fails the sovereignty test. The "Expert" mode’s router either sends even trivial tasks directly to the GLM-based model rather than Yandex’s own proprietary system, or Yandex’s supposedly "sovereign model" is itself built upon a Chinese foundation. It appears that the Russian Alice AI LLM is simply not being used in the assistant.
