Baidu's ERNIE-5.0-Preview-1120 Debuts on LMSYS Arena Vision Leaderboard, Tops Domestic Models
BEIJING — Baidu has released ERNIE-5.0-Preview-1120 for public testing on LMSYS Chatbot Arena, where the new multimodal model immediately secured a spot in the global Top 15 on the Vision leaderboard with a score of 1206. The preview ranks first among domestic Chinese models and performs on par with leading frontier systems including Claude Sonnet 4 and GPT-5-high.
The announcement marks the latest milestone for Baidu’s ERNIE 5.0 series, which the company describes as a 2.4 trillion-parameter unified multimodal model trained from scratch. Unlike late-fusion architectures that process different modalities separately, ERNIE-5.0 integrates text, image, video, and audio within a single autoregressive framework.
According to Baidu’s official ERNIE Blog, “Our newly released ERNIE-5.0-Preview-1120 has entered the LMArena Vision Leaderboard for the very first time! It lands straight in the Top 15 with a score of 1206, ranked top 1 in domestic models, on par with Claude Sonnet 4 and GPT-5-high!”
Technical Details and Benchmark Performance
ERNIE-5.0-Preview-1120 is now available for direct side-by-side testing against other top models at lmarena.ai. The Vision leaderboard evaluates multimodal understanding and reasoning capabilities through crowdsourced human votes, making the ranking a strong indicator of real-world perceptual performance.
This release follows closely on the heels of ERNIE-5.0-Preview-1022, which achieved a tied No. 2 global position on the LMSYS Text leaderboard. The rapid iteration highlights Baidu’s aggressive development pace in both language-only and multimodal domains as Chinese AI labs intensify competition with U.S. leaders.
The model forms part of a broader ERNIE 5.0 family that includes both large language models and vision-language models. Earlier reports noted that certain ERNIE 4.5 variants had outperformed models such as DeepSeek-V3 and Qwen-235B on specific benchmarks while remaining competitive with OpenAI’s o1 reasoning model, though the 5.0-Preview-1120 Vision results represent its first official placement on the Arena Vision board.
Industry Context
Baidu’s ERNIE series has long been one of China’s flagship large model families, evolving from early language models to today’s massive unified multimodal systems. The 2.4 trillion parameter scale places ERNIE-5.0 among the largest models publicly discussed by major Chinese labs, signaling substantial compute investment.
The decision to prioritize LMSYS Chatbot Arena testing reflects the growing importance of community-driven, Elo-based leaderboards in assessing frontier AI capabilities. These blind comparison platforms have become de facto benchmarks because they measure human preference rather than academic metrics alone.
Impact on Developers and the AI Ecosystem
For developers, the availability of ERNIE-5.0-Preview-1120 on Arena provides an immediate opportunity to evaluate its multimodal reasoning against global competitors without needing API access or local inference resources. The model’s strong domestic ranking may encourage greater adoption within China’s AI application ecosystem, particularly in areas requiring sophisticated image, video, and audio understanding.
The release also intensifies the global race in unified multimodal architectures. While many Western labs have pursued separate specialist models or relatively simpler fusion techniques, Baidu is betting on end-to-end autoregressive training across all modalities from the ground up.
What's Next
Baidu has not yet disclosed a full technical report, training details, or commercial API timeline for the complete ERNIE-5.0 model. The current preview appears focused on gathering community feedback through Arena battles before wider deployment.
Industry observers expect Baidu to continue iterating on both the text and vision leaderboards in coming weeks, potentially releasing additional preview versions. The company has a track record of rapidly incorporating Arena test results into subsequent model improvements.
Users can begin testing ERNIE-5.0-Preview-1120 immediately at lmarena.ai. Further details are available on the official ERNIE Blog.
The development underscores China’s continued push to close the perceived gap with U.S. frontier labs in both raw scale and multimodal integration, with real-world human preference benchmarks serving as the primary measuring stick.