Telegram's AI assistant, the one baked into the messaging app that hundreds of millions of people use daily, is running Alibaba's Qwen 3.5 under the hood.
A researcher confirmed this by extracting the system prompt from Telegram's AI feature through a straightforward prompt injection technique - basically asking the model to reveal its own instructions. The model identified itself as Qwen 3.5, a large language model developed by Alibaba Cloud. The extracted system prompt also revealed specific behavioral guidelines Telegram configured, including instructions about tone, content restrictions, and how the assistant should handle various query types.
This matters for a practical reason: when you use an AI feature inside an app, you rarely know what model is actually processing your messages. Telegram never publicly disclosed which model powers its AI assistant, and most users probably assumed it was either custom-built or running on one of the big Western models like GPT-4 or Claude. Instead, it's Qwen - a strong open-weight model family, but one that comes with different data handling practices and training data considerations than what many Western users might expect.
The broader pattern here is worth paying attention to. As more apps bolt on AI features, the actual model behind the interface is increasingly obscured. Your favorite productivity app might say "AI-powered" without telling you whether that means GPT-4o, Claude, Llama, Qwen, or something else entirely. Each of those models has different strengths, weaknesses, and privacy implications. Telegram choosing Qwen 3.5 likely came down to cost and licensing flexibility since Qwen's open-weight models can be self-hosted, but users deserve to know what's processing their conversations.