ERNIE-5.0-Preview-1220 Becomes Sole Chinese Model in LMArena Vision Top 10
BEIJING — Baidu’s ERNIE-5.0-Preview-1220 achieved a score of 1,226 on LMSYS Arena’s Vision Arena, securing eighth place globally and becoming the only Chinese model to crack the top 10, the company announced Wednesday.
The milestone was recorded when LMSYS released its latest Vision Arena rankings on Jan. 8. ERNIE-5.0-Preview-1220 now ranks as China’s top-performing vision model and the sole domestic entry among the global leaderboard’s elite tier, ahead of several high-profile Western and other Chinese multimodal systems.
According to Baidu’s official ERNIE Blog, the preview model outperformed Google’s Gemini-2.5-Flash and OpenAI’s GPT-5-Chat on vision tasks while placing second globally in creative writing, trailing only Gemini-3-Pro. The result underscores growing competitiveness from Chinese labs in multimodal AI evaluation, an area long dominated by U.S. frontier models.
Strong Multimodal Showing
LMArena’s Vision Arena uses crowdsourced human preference votes to rank models on visual understanding capabilities, including image reasoning, chart interpretation, and creative visual tasks. ERNIE-5.0-Preview-1220’s 1,226-point score positioned it ahead of multiple competing systems, including Qwen3-VL-235B-A22B-Instruct, according to community discussion on X.
The achievement follows Baidu’s earlier success on the text-only leaderboard. In a previous update, ERNIE-5.0-0110 scored 1,460 points to reach eighth place globally on LMArena’s primary text ranking, also becoming the only Chinese model in that top 10 at the time and surpassing both GPT-5.1-High and Gemini-2.5-Pro.
Baidu has positioned ERNIE 5 as a major upgrade to its foundational large language model family, with particular emphasis on enhanced reasoning, mathematical capabilities, and multimodal integration. The “Preview-1220” designation indicates an early preview build released in December 2020 nomenclature, though the exact parameter count and training details have not been publicly disclosed in the announcement.
Industry Context
The result arrives amid intensifying global competition in multimodal AI. While U.S. labs such as OpenAI and Google continue to lead many English-centric and general capability benchmarks, Chinese models have made consistent gains on independent arenas like LMSYS, particularly as evaluation methodologies increasingly incorporate non-English and culturally diverse prompts.
Being the sole Chinese representative in the Vision Arena top 10 highlights both Baidu’s progress and the relative scarcity of other domestic models in the highest echelons of this particular leaderboard. Several other major Chinese labs, including Alibaba’s Qwen team, have released competitive vision-language models, yet none currently appear in the global top 10 alongside ERNIE-5.0-Preview-1220.
Impact on Developers and Users
For developers building applications that require strong visual understanding, the ranking provides an independent signal that Baidu’s latest model family may be worth evaluating, especially for use cases involving Chinese-language content, cultural context, or mixed text-vision workflows. The model’s strong creative writing ranking also suggests potential utility in generative tasks that blend textual creativity with visual inputs.
Enterprise users in China may find the result particularly relevant as domestic AI infrastructure and data sovereignty requirements continue to favor locally developed models. The dual strength in vision and creative writing positions ERNIE 5 as a versatile multimodal contender.
What’s Next
Baidu has not yet announced a full production release timeline for ERNIE-5.0 or detailed availability through its ERNIE Bot platform and API services. The “Preview” designation suggests further iterations and potential performance improvements are likely before a stable version ships.
LMSYS Arena rankings are dynamic and subject to change as new models enter the arena and additional votes are collected. Community observers expect other Chinese labs to challenge ERNIE’s current position in the coming weeks.
Baidu said it will continue releasing updates on ERNIE model performance across both text and vision arenas. The company has not disclosed whether the 1220 preview will be made publicly accessible for direct testing on the LMSYS Chatbot Arena in the immediate term.
The article is based on Baidu’s official ERNIE Blog announcement and verified community reports from Jan. 8, 2025.