DeepSeek Releases V4-Flash-Vision-Exp Model

DeepSeek has expanded its API platform with the release of DeepSeek-V4-Flash-Vision-Exp, an experimental multimodal version of its lightweight V4-Flash model. This new release retains the text capabilities of the base model, including its reasoning, agentic behavior, and world knowledge, while adding native image processing. Under the hood, the model uses an efficiency-optimized Mixture-of-Experts architecture containing 284 billion total parameters and 13 billion activated parameters. It supports a one-million-token context window and utilizes hybrid attention for processing long contexts.

According to DeepSeek, the model approaches the performance of Anthropic's Claude Opus 4.8 on multimodal agent tasks. In internal evaluations using its Harness Minimal Mode, the model outperformed Opus 4.8 on three of eleven benchmarks: DeepSWE by 1.3 points, Agents' Last Exam by 1.6 points, and ZeroBench by 1.0 point. However, it trailed on the remaining eight benchmarks. Notably, it fell 12 points behind on NL2Repo, scoring 57.7 compared to Opus 4.8's 69.7. On Terminal Bench 2.1, the model scored 83.9, beating the base V4-Flash score of 82.7 but trailing Opus 4.8's 85.0. It also lagged behind Opus 4.8 by roughly eight points on DSBench-Hard.

The OpenAI-compatible API allows developers to input images via base64 data URLs up to 48 megabytes, standard URLs, or a new free Files API that supports uploads up to 64 megabytes. Images are resized to fit between 384-by-384 and 800-by-800 pixels, capping consumption at 384 tokens per image. Image inputs are billed at standard V4-Flash text rates, which cost $0.14 per million cache-miss input tokens, $0.0028 per million cached input tokens, and $0.28 per million output tokens, with a maximum output of 384,000 tokens.

For developers, this release simplifies agentic workflows that require visual context, such as GUI automation, chart analysis, and document triage. Instead of routing visual tasks to a separate, expensive model, practitioners can now handle text and vision within a single, low-cost API call. The free Files API also allows developers to upload an image once and reuse its identifier across multiple requests, reducing bandwidth costs for iterative agent loops.