DeepSeek Launches Free Multimodal AI Model to Challenge Claude Opus 4.8

Chinese artificial intelligence startup DeepSeek has released DeepSeek-V4-Flash-Vision-Exp, an experimental multimodal model that adds image and screenshot interpretation directly into its lightweight V4 Flash architecture. The launch intensifies competitive pressure on Western AI developers following the recent releases of Kimi K3 and Qwen 3.8, positioning DeepSeek to challenge high-tier models like Anthropic’s Claude Opus 4.8.

While rival offerings such as Claude and ChatGPT require premium pricing tiers for multimodal features, DeepSeek maintains its base text-only rate for the new visual capabilities. Company documentation indicates the model retains the underlying reasoning, broad domain knowledge, and autonomous agent capabilities of V4 Flash, enabling unsupervised multi-step tasks across visual and textual data — a foundation the API now extends to visual workflows.

Those agent capabilities now reach developers through an API that accepts mixed text and visual inputs via base64 encodings, external URLs, or a dedicated File API. The File API assigns unique identifiers to uploaded images, streamlining multi-turn agent workflows by eliminating repeated uploads.

Each uploaded file is tokenized at a fixed cap of 384 tokens, a limit that shapes the system’s batch capacity of up to 600 images per request across JPEG, PNG, GIF, and WebP formats. DeepSeek enforces strict file limits of 32 MiB for direct payloads and 64 MiB through its File API, automatically scaling smaller images up to 384 by 384 pixels and larger files down to an equivalent resolution of 800 by 800 pixels.

The 800-by-800-pixel ceiling on scaled images defines the practical resolution ceiling for the deepseek-v4-flash-vision-exp endpoint, which developers invoke after generating API keys through the DeepSeek developer portal. The platform delivers combined visual recognition and structured reasoning outputs within a single API response call.