Baidu's ERNIE-5.0-Preview-1022 Climbs to #2 on LMSYS Arena Text Leaderboard
BEIJING — Baidu has released ERNIE-5.0-Preview-1022 for public testing on LMSYS Chatbot Arena (LMArena), where the model now ranks #2 globally on the Text leaderboard with a score of 1432, according to the company's official announcement.
The preview model achieved the #2 position (tied) on the widely respected crowdsourced benchmark that evaluates large language models through real-world user votes. It also holds the #1 spot in China. Baidu said the model excels in creative writing, complex reasoning, and instruction following, areas where it reportedly outperforms several leading competitors in head-to-head Arena matchups. The full ERNIE 5.0 model is scheduled for official unveiling at Baidu World 2025 on November 13.
Technical Details and Performance
According to Baidu's ERNIE Blog, ERNIE-5.0-Preview-1022 is a natively omni-modal mixture-of-experts (MoE) model with 2.4 trillion parameters. The architecture processes text, images, audio, and video within a unified model while activating fewer than 3% of parameters per inference step. This efficient activation strategy reportedly enables strong performance with reduced computational overhead.
The model has been evaluated across more than 40 authoritative benchmarks. Baidu claims it surpasses previous domestic leaders including DeepSeek V3 and Qwen2.5-235B in multiple categories while remaining competitive with top international models. On the LMArena Text leaderboard, its 1432 Elo score places it immediately behind the current global leader and ahead of several prominent Western models.
Independent community discussion on platforms including X and Reddit has highlighted the model's reported strengths in creative tasks and complex reasoning chains. Some observers noted it outperforming "GPT-5-high" in certain Arena categories, though these specific matchups have not been independently verified by third-party organizations.
Context Within China's AI Landscape
Baidu's release continues the rapid pace of innovation from Chinese AI labs. The ERNIE series has long served as the company's flagship large language model family, powering its search engine, cloud services, and consumer applications. With ERNIE-5.0-Preview-1022, Baidu appears to be pushing toward unified multimodal understanding rather than relying on separate specialist models.
The timing is notable as multiple Chinese organizations have recently demonstrated competitive or leading performance on global benchmarks. The LMArena leaderboard, maintained by LMSYS Org, has become one of the most watched independent evaluations because it reflects blind user preferences rather than academic benchmark scores alone.
Impact on Developers and the Industry
For developers, the availability of ERNIE-5.0-Preview-1022 on LMArena provides an opportunity to test the model directly against other frontier systems through anonymous battles. The model's strong showing in creative writing and instruction following could make it particularly attractive for content generation, agentic workflows, and complex reasoning applications.
The efficient MoE design, which activates only a small fraction of the 2.4T parameters, suggests potential cost advantages for inference at scale. However, actual production pricing and API availability details have not yet been disclosed by Baidu.
Industry observers see the release as further evidence that Chinese labs are closing the gap — and in some cases challenging — Western leaders on both performance and multimodal capabilities. The upcoming full launch at Baidu World 2025 is expected to include additional technical details, benchmark breakdowns, and likely commercial availability information.
What's Next
Baidu has confirmed that the complete ERNIE 5.0 model will be revealed on November 13 at Baidu World 2025. The company has not yet published a detailed technical report or comprehensive benchmark suite for independent verification.
Developers and researchers can currently interact with the preview version on the LMArena platform to test its capabilities in real-time battles. Further updates on model weights, API access, and enterprise offerings are expected to follow the November 13 event.
The rapid iteration of the ERNIE series underscores the intense competition in the global foundation model space, where both performance on public leaderboards and practical multimodal capabilities have become key differentiators.
This article is based on Baidu's official ERNIE Blog announcement and related community reporting. Some comparative performance claims remain unverified by independent third parties.