DeepSeek Visual Model Multimodal Agent Approaches Opus 4.8! A 2:2 Tie Across 4 Metrics

Dongcha Beating AI Newsflash: DeepSeek has released the first batch of Agent scores for V4-Flash-Vision-Exp. In 4 multimodal Agent evaluations, it tied with Opus 4.8 with 2 wins and 2 losses: Agents' Last Exam scored 27.3 vs. 25.7, ZeroBench scored 35.0 vs. 34.0; ApexBench scored 36.5 vs. 39.4, and Chartography scored 64.3 vs. 65.0. Compared to the text-only version V4-Flash-0731, the visual version improved ApexBench from 26.2 to 36.5 and Agents' Last Exam from 25.2 to 27.3. However, since these two evaluations inherently include images, the official note clarifies that V4-Flash directly ignores multimodal content in these cases. Therefore, the results mainly reflect the added capability of visual input rather than a sudden improvement in text-only reasoning. With the addition of visual input, the text Agent's performance did not show significant decline either. Among 7 text-based evaluations, Vision Exp outperformed V4-Flash-0731 in 6 of them. For instance, DeepSWE improved from 54.4 to 59.3, even surpassing Opus 4.8's 58.0; Toolathlon scored 75.9, nearly matching Opus 4.8's 76.2. It is important to note that these results are from DeepSeek's official internal testing and not from third-party independent leaderboards. The DeepSeek series uses the DeepSeek Harness in minimalist mode for public Code Agent text tasks, with the inference level set to max. [Original Link]