DeepSeek Launches V4-Flash-Vision-Exp, Outperforming Claude in Application Tests

DeepSeek launched the V4-Flash-Vision-Exp model, a multimodal version built on the MoE architecture with 284 billion parameters, outperforming the Claude Opus 4.8 model in the ALE benchmark consisting of over one thousand application tasks, as well as in the ZeroBench test.

The model is distinguished by its use of HCA and CSA compression techniques, which reduced the cost of processing a one-million-token context by 73%, making it an attractive economical choice for applications requiring long contexts.

This launch strengthens DeepSeek's position in the large language model market, as it continues to compete at the top of international benchmarks with models that combine computational efficiency and high performance simultaneously.

This model poses a direct challenge to competitors, as it combines performance superiority over major global models with a significant reduction in cost—a combination that redraws developer expectations for multimodal AI models.

What do these terms mean?

MoE Architecture (Mixture of Experts): An architecture that activates only a fraction of the model's parameters for each task, reducing computation without compromising overall capability.

ALE Benchmark: An evaluation platform that simulates over one thousand real application tasks to measure models' ability to interact with real-world environments.

HCA/CSA Compression: Two compression techniques that reduce the data processed in long contexts, fundamentally lowering computational costs.