DeepSeek Ships Its First Vision Model: V4-Flash-Vision-Exp

DeepSeek quietly added its first vision-capable model to the API on August 21, 2026: an experimental release called DeepSeek-V4-Flash-Vision-Exp, accessible via model='deepseek-v4-flash-vision-exp'. It’s a multimodal variant of the V4-Flash line that keeps the text model’s agent and reasoning performance intact while adding real image-understanding ability.

Details

• Text performance holds steady: on pure text tasks, V4-Flash-Vision-Exp scores on par with the standard DeepSeek-V4-Flash across agent, reasoning, and knowledge benchmarks — Terminal Bench 2.1 (83.9), NL2Repo (57.7), DeepSWE (59.3), and DSBench-Hard (63.6)
• New visual benchmarks: the model posts 25.7 on AutomationBench (Public), 36.5 on ApexBench (Pass@1), 27.3 on Agents’ Last Exam, 64.3 on Chartography, and 35.0 on ZeroBench (Pass@5)
• Big jump over the text-only model: DeepSeek describes the visual-understanding gain as a “significant leap” over standard V4-Flash, bringing multimodal agent capabilities closer to Anthropic’s Opus-4.8
• Experimental status: the model is explicitly labeled experimental, suggesting further iteration before a stable, non-”exp” release
• API-only for now: available through the DeepSeek API; no changes were announced yet for the consumer app or web interface

What happened next

DeepSeek has not announced pricing specific to the vision variant separately from the existing V4-Flash tiers, nor a firm timeline for graduating it out of experimental status. Developers can start testing multimodal agent workflows — chart reading, screenshot-based automation, and visual QA — immediately via the API using DeepSeek’s vision guide documentation.