-+ 0.00%
-+ 0.00%
-+ 0.00%
YunZhisheng (09678) released a new upgraded version of the big model U2-Flash-1118
Share
Listen to the news

Zhitong Finance App News, Yun Zhisheng (09678) issued an announcement, and the company officially released U2-Flash-1118. This model is a new upgraded version of U2-Flash, which has achieved a comprehensive leap forward in the three major capabilities of complex programming, automatic agent execution, and multi-modal understanding. In many authoritative evaluations, U2-Flash-1118 performed ahead, achieving excellent results in the three major fields of complex programming and agent execution, native visual understanding, and scientific and mathematical reasoning. The specific evaluation results are as follows:

In terms of native visual understanding, the model ranked in the top 2 of the 3 visual rankings that participated in the review. Among them, the OCRBench-V2 English and Chinese evaluation scores surpassed DeepSEEK-v4.1-Flash-1017 and GLM 5.3-Flash, showing excellent graphic detail recognition, spatial relationship perception, and comprehension of Chinese and English document layouts.

In terms of complex programming and agent execution, the model improved compared to U2-Flash in all 4 engineering evaluations participating in the review. Among them, the DeepSWE v1.1 evaluation score ranked second, surpassing MiniMax-M3.1-Flash and Qwen3.8-Max-0902, showing excellent large-scale code warehouse analysis, complex engineering requirements dismantling, multi-round code modification, and terminal task execution capabilities.

In terms of scientific and mathematical reasoning, the model ranked first and second respectively in the 2 assessment of reasoning ability that participated in the evaluation. Among them, GPQA-Diamond ranked first in the evaluation score, surpassing Kimi K3 and Qwen3.8 Max, showing excellent mathematical reasoning ability for difficult scientific questions and answers and contests.

On the basis of keeping the scale of the original language base parameters unchanged, U2-Flash-1118 adds a visual module with about 1B parameters to enhance native multi-modal understanding ability. At the same time, through post-training and Agent Harness collaborative optimization, the model's ability to perform complex tasks and token usage efficiency are further improved. Its ability upgrades are mainly in the following three areas:

(1) Enhanced multi-modal understanding: The model introduces native image input capabilities, can directly understand UI screenshots, design drafts, and document layouts, and supports graphic Q&A, interface understanding, and assisted development based on screenshots.

(2) Optimize complex task execution: Through the recursive self-improvement (RSI) mechanism, task execution feedback is transformed into targeted training tasks, and the model and agent execution framework are continuously optimized through error correction training, independent testing, and ability regression verification, and the complete development process from requirement understanding, problem location, code modification to test verification can be independently promoted.

(3) Token efficiency improvement: In terms of agent execution efficiency, repeated exploration, redundant tool calls, and context handling during task execution are reduced through execution trajectory optimization, key action refinement, dynamic budget allocation, and Agent Harness collaborative optimization. Under the same task set and consistent evaluation conditions, compared with the original U2-Flash version, the total token consumption was reduced by about 37.0%, and the number of model calls was reduced by about 44.8%, further reducing task execution costs while maintaining the quality of task completion.

U2-Flash-1118 further enhances the comprehensive ability of the model in actual application scenarios through enhanced multi-modal understanding ability, improved ability to execute complex tasks, and optimized agent execution efficiency, and provides more efficient model services for developers and enterprise users.

Disclaimer:Webull uses external vendor Google Translation Service for news translations where we endeavour to ensure these are correct, however, we recommend that you please double-check this information accordingly. Webull is not responsible for translation errors or issues.
What's Trending