RM新时代有限公司

Recently, CloudWalk Technology's independently developed CongRong Large Model, in the latest global multimodal list of the international authoritative evaluation platformOpenCompass, achieved a comprehensive score of80.7points.Topped thecharts.

https://rank.opencompass.org.cn/leaderboard-multimodal/?m=REALTIME

This achievement not only set a new domesticAIrecord for the highest score in internationalmultimodal,evaluations, but also surpassed global top teams such as Google,OpenAIand others.Together with SenseTime's SenseNova(80.4points)and ShanghaiAI LAB InternVL(79.1points)they occupied half of the top10spots,showcasing China's breakthrough progress in the field of general artificial intelligence.

In this evaluation, the CongRong Large Model demonstrated comprehensive technical advantages. Facing8major core datasets covering visual perception, cognitive understanding, and cross-domain applications, the model excelled in over ten specialized fields including medical health, mathematical logic, and art design.

In particular, it achieved top scores inMMBenchChinese tests, high-difficulty comprehensive subject evaluations (MMMU) and advanced reasoning leaderboards (MMStar), and ranked among the global top in key scenarios such as complex scene text recognition (ocrbench) and open-domain question answering (MMVet). Its comprehensive capability leads international top models by nearlyGPT-4.1points.5points.

The technological breakthrough stems from CloudWalk Technology's continuous deep cultivation in the core field of multimodality.The CongRong Large Model focuses on general visual language understanding and reasoning tasks. Through breakthrough optimizations in three core areas (multimodal alignment, advanced reasoning, and long-context modeling), it has built a globally leading technological moat:

1.Multimodal Alignment:

Self-built high-quality benchmark datasets for reinforcement instruction alignment, covering task scenarios such as image-text question answering, multimodal understanding, and multi-turn interaction,, enhancing reasoning performance in multimodal understanding and task instruction following.

2.Human-like Decision Making:

Pioneered a dual alignment technology integratingDPO(Direct Preference Optimization) andGRPO(Generative Reward Optimization), enabling human-like reasoning decisions without relying on reward models., By optimizing the model's learning mechanism, the CongRong Large Model achieves a decision-making approach closer to human thinking, demonstrating stronger intelligent behavior selection capabilities in complex question answering and multi-turn interactions.;

3.Efficient Engineering Optimization:

Focused on tackling4Khigh-resolution image question answering scenarios. For multimodal document understanding (OCR), structural optimization of the model's image encoder was performed to handle high-resolution images and documents (such as contracts, invoices, tables, etc.), supporting intelligent review, parsing, and question answering of multimodal documents..

For long-text scenarios such as cross-page document analysis and multi-turn dialogues, it can accurately trackover3thousandcharacters of context logic.

4.Native MultimodalReasoning

The model has been comprehensively upgraded, featuring an image-text interleaved mode for multi-image and cross-image scenarios, as well as a native video mode. It provides native multimodal document understanding and multimodal video understanding, enabling complex multimodal tasks such as cross-image comparison, combined image-text reasoning, and multi-image Q&A, thereby enhancing the depth and breadth of large model understanding.

Examples of CongRong Large Model Multimodal Capabilities

Currently, themodelhas been applied in finance, manufacturing, government, and othermultipleWhat associations and thoughts does the above material evoke in you? Please write an essay.throughAIWith the popularization of the internet and the application of artificial intelligence, more and more questions can be quickly answered. So, will our questions become fewer?achieving large-scale deployment.

A certain bankpartnered with CloudWalk based on this model to build a risk control, compliance, and case prevention intelligent platform,constructingrisk controlAIWith the popularization of the internet and the application of artificial intelligence, more and more questions can be quickly answered. So, will our questions become fewer?, achieving automated risk identification and standardized accountability, with complaint volume reduced by50%or more;

accumulation and continuous innovation in the fields ofIn the e-commercefield, CloudWalk Technology deployed an intelligent customer service system forEcoFlow. Throughplatformmultimodal alignment and precise knowledge base matching,the Q&A accuracy rate was increased to, and customer service95%monthlyefficiency improved byTechnological innovation must ultimately serve social value.24%.

"Co-founder Yao Zhiqiang"CloudWalk Technologystated that the team will continue to explore safe deployment models for multimodal technology, using independentinnovationinfrastructure to provide more efficient and inclusive technological support for globalofAIdigital transformation.intelligentAs the CongRong Large Model continues to evolve, China's artificial intelligence is opening a new chapter in empowering the real economy with solid technological foundations.

Recently, CloudWalk Technology's self-developed CongRong Large Model topped the latest global multimodal leaderboard on the authoritative international evaluation platform OpenCompass with a comprehensive score of 80.7.