UNISOUND (09678) releases a fully upgraded large model version U2-Flash-1118.

date
08:12 18/11/2026
avatar
GMT Eight
UNISOUND (09678) announced that the company has officially released U2-Flash-1118. This model is a brand-new upgraded version of U2-Flash, achieving comprehensive leaps in three major capabilities: complex programming, Agent autonomous execution, and multimodal understanding. In multiple authoritative evaluations, U2-Flash-1118 has delivered leading performance, achieving excellent results in three major areas: complex programming and Agent execution, native visual understanding, and scientific and mathematical reasoning. The specific evaluation results are as follows:
UNISOUND (09678) announces that the company has officially released U2-Flash-1118. This model is a brand-new upgraded version of U2-Flash, achieving comprehensive leaps in three major capabilities: complex programming, autonomous Agent execution, and multimodal understanding. In multiple authoritative evaluations, U2-Flash-1118 has delivered leading performance, achieving excellent results across three major domains: 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 two across all 3 visual benchmarks it participated in. Among them, its scores on both the English and Chinese evaluations of OCRBench-V2 exceeded those of DeepSeek-V4.1-Flash-1017 and GLM 5.3-Flash, demonstrating outstanding capabilities in graphical detail recognition, spatial relationship perception, and Chinese and English document layout understanding. In terms of complex programming and Agent execution, the model improved over U2-Flash across all 4 engineering evaluations it participated in. Among them, it ranked second on the DeepSWE v1.1 evaluation, surpassing MiniMax-M3.1-Flash and Qwen3.8-Max-0902, demonstrating excellent capabilities in large code repository parsing, decomposition of complex engineering requirements, multi-round code modification, and terminal task execution. In terms of scientific and mathematical reasoning, the model ranked first and second respectively in the 2 reasoning capability evaluations it participated in. Among them, it ranked first on the GPQA-Diamond evaluation, surpassing Kimi K3 and Qwen3.8 Max, demonstrating outstanding high-difficulty scientific question answering and competition-level mathematical reasoning capabilities. U2-Flash-1118, while keeping the parameter scale of the original language foundation unchanged, adds a visual module of approximately 1B parameters to enhance native multimodal understanding capabilities. At the same time, through post-training and Agent Harness collaborative optimization, it further improves the model's complex task execution capabilities and Token usage efficiency. Its capability upgrades are mainly in the following three aspects: (1) Enhanced multimodal understanding capabilities: The model introduces native image input capabilities, can directly understand UI screenshots, design mockups, and document layouts, and supports image-text Q&A, interface understanding, and screenshot-based assisted development. (2) Optimized complex task execution: Through a recursive self-improvement (RSI) mechanism, task execution feedback is converted into targeted training tasks, and through error-correction training, independent testing, and capability regression validation, the model and Agent execution framework are continuously optimized, enabling autonomous advancement of the complete development process from requirement understanding and problem localization to code modification and test validation. (3) Improved Token efficiency: In terms of Agent execution efficiency, through execution trajectory optimization, key action extraction, dynamic budget allocation, and Agent Harness collaborative optimization, repetitive exploration, redundant tool calls, and context transfer during task execution are reduced. Under the same task set and consistent evaluation conditions, compared with the original version U2-Flash, total Token consumption decreased by approximately 37.0%, and the number of model calls decreased by approximately 44.8%, further reducing task execution costs while maintaining task completion quality. U2-Flash-1118 further enhances the model's comprehensive capabilities in real-world application scenarios through enhanced multimodal understanding, improved complex task execution, and optimized Agent execution efficiency, providing more efficient model services for developers and enterprise users.