DeepSeek Launches Experimental Multimodal Model to Challenge Anthropic
DeepSeek has released an experimental multimodal model, claiming it rivals Anthropic's Opus-4.8 in performance. According to DeepSeek, their model, named DeepSeek-V4-Flash-Vision-Exp, outperforms Opus-4.8 in three out of eleven benchmarks, winning on DeepSWE, Agents' Last Exam, and ZeroBench.
DeepSeek-V4-Pro: A New Release
DeepSeek-V4-Pro was launched on August 21, 2026, featuring an experimental multimodal model with enhanced capabilities. The company's API platform now includes DeepSeek-V4-Flash-Vision-Exp, which combines text-only capabilities with the ability to interpret images and screenshots.
Benchmarks and Performance
DeepSeek's published table reveals mixed results. While the new model excels in certain benchmarks, it trails behind on eight others, with significant gaps on NL2Repo and DSBench-Hard. The closest scores are in Toolathlon-Verified, Chartography, and Terminal Bench 2.1, where the difference is minimal.
Multimodal Agent Performance
A notable claim is the significant improvement in multimodal agent performance over the previous V4-Flash model. On ApexBench and Agents' Last Exam, the new model demonstrates substantial jumps, but DeepSeek notes that the older model ignores multimodal elements, making a direct comparison biased.
Underselling a Result
DeepSeek's disclosure of the vision mode's performance is commendable, yet the company seems to undersell one significant result. The table does not include a comparison with Anthropic's latest Opus 5 model, leaving an unknown gap in the evaluation.