Baidu's ERNIE-5.0-Preview-1203 Tops LMArena Text Rankings
BEIJING — Baidu’s ERNIE-5.0-Preview-1203 has claimed the top spot among Chinese models on the influential LMSYS Chatbot Arena (LMArena) leaderboard, scoring 1,451 points and surpassing several prominent international competitors.
According to the latest LMArena rankings released Thursday, the preview model not only became the highest-ranked text model from China but also outperformed well-known entries including Claude-Opus-4-1 and GPT-5.2. The achievement was announced on Baidu’s official ERNIE Blog and shared across platforms including Hugging Face.
LMArena, the community-driven blind-testing platform operated by LMSYS, ranks large language models based on human preference votes across millions of pairwise comparisons. A score of 1,451 places ERNIE-5.0-Preview-1203 at the head of the current text-only leaderboard for Chinese developers, marking a significant milestone for Baidu’s ongoing ERNIE series.
The model is described by Baidu as merging creativity with advanced reasoning capabilities. Developers can now experience ERNIE-5.0-Preview-1203 directly on the LMArena platform, with the company stating that the release opens “a new era of language intelligence.”
This marks the latest in a string of strong showings for the ERNIE-5.0 preview family. A related vision model, ERNIE-5.0-Preview-1220, previously achieved a score of 1,226 on LMArena’s Vision Arena, securing first place among Chinese models and eighth globally in multimodal rankings.
Baidu has positioned ERNIE as its flagship large language model family, competing in both domestic and international arenas against offerings from OpenAI, Anthropic, Google and other Chinese labs including Alibaba, Moonshot AI and DeepSeek. The 1,451-point score represents a notable jump from earlier ERNIE iterations and signals accelerating progress in Baidu’s foundational model development.
Industry Impact
For the Chinese AI ecosystem, the result is likely to be viewed as validation of continued domestic investment in large language models despite export controls on advanced hardware. Baidu, one of China’s earliest AI pioneers through its Apollo autonomous driving project and ERNIE knowledge-enhanced models, has steadily pushed the ERNIE series toward frontier performance.
Developers and researchers can now test the model on LMArena’s public arena, providing immediate feedback through the platform’s Elo-based ranking system. The availability of a preview version on the arena also allows Baidu to gather large-scale human preference data that can be used to further refine subsequent releases.
The news arrives as global attention on Chinese AI capabilities continues to grow, with several domestic models recently demonstrating competitive performance on public benchmarks despite constraints on training compute.
What’s Next
Baidu has not yet disclosed full technical specifications for ERNIE-5.0-Preview-1203, including parameter count, training data details or context window size. The “Preview-1203” naming convention suggests an internal snapshot dated December 3, indicating the model may still be undergoing rapid iteration.
The company is expected to provide additional details in coming weeks, potentially including API availability, enterprise pricing and integration options within Baidu’s cloud platform. Further gains on the overall LMArena leaderboard, rather than just the Chinese ranking, would position ERNIE-5.0 as a more direct global competitor to leading models from OpenAI and Anthropic.
LMArena typically updates its rankings multiple times per week as new models are submitted and additional votes are collected. Continued monitoring of the leaderboard will reveal whether ERNIE-5.0-Preview-1203 can maintain or improve its position as more users interact with the model.
Baidu says it will continue releasing updates to the ERNIE-5.0 series, with the current LMArena appearance serving as an early public preview of capabilities expected in the full ERNIE 5.0 release.