- Published on
Bilibili launches 'AI Infinite Arena' LLM evaluation leaderboard: claims '100 global models compete together', GPT-6 ...
- Authors
- Name
- DeepHorus
On September 20, Bilibili announced the launch of the "AI Infinite Arena" large model evaluation leaderboard, and simultaneously released the first round of model rankings. The leaderboard shows that GPT-6 Astra took the top spot in evaluations by 10 UP主 (content creators), defeating GLM-5.3 to win the most first-place finishes; among the top five, domestic large models occupied three seats. The leaderboard URL has now been made public, and users can visit Bilibili to view real-time rankings.
According to Bilibili, the "AI Infinite Arena" is a competitive arena that gathers AI large model evaluations from Bilibili UP主. It consists of real-world scenario tests of hundreds of large models conducted by UP主 from various fields, with evaluation topics covering code, reasoning, collaboration, knowledge, and more. Unlike traditional benchmark tests, this arena does not impose restrictions on topics or evaluation dimensions. UP主 from different sections set their own topics based on real workflows, professional applications, fun ideas, etc., and in same-topic PKs, the actual performance differences between models are intuitively presented.
The leaderboard covers comparative evaluations of mainstream models including DeepSeek, Kimi, ChatGPT, Claude, Gemini, Doubao, Qianwen, Hy, MiniMax, and others. Rankings are updated in real time, and registration remains open to all UP主 across the platform. This means any UP主 can participate in evaluations, continuously enriching the leaderboard content.
Public information shows that over the past year, Bilibili's AI knowledge content consumption time increased by 72% year-on-year, and the average monthly number of users watching AI content exceeded 190 million. The platform has seen a large amount of AI content covering release, discussion, evaluation, usage, help-seeking, co-construction, and other stages, providing a content foundation for the "AI Infinite Arena." Bilibili stated that the leaderboard aims to help users more intuitively understand the performance of large models in real-world scenarios.
As more UP主 participate in evaluations, the leaderboard rankings will continue to change, and the competitive landscape between domestic and overseas models deserves attention. Industry insiders believe that such crowd-sourced leaderboards based on real scenarios are expected to become an important reference alongside traditional benchmark tests, and also provide user-side feedback to model developers. In the future, the richness and participation scale of AI evaluation content will affect the reference value of this leaderboard.
