DeepSeek Launches Experimental Multimodal Vision Model DeepSeek‑V4‑Flash‑Vision‑Exp

DeepSeek announced that the experimental multimodal model DeepSeek‑V4‑Flash‑Vision‑Exp is now available on the DeepSeek API platform. Developers can invoke it by setting the model parameter to deepseek-v4-flash-vision-exp. The "Exp" suffix indicates that this is an experimental version rather than the official multimodal release of V4‑Flash.

Two hours earlier the DeepSeek Harness repository was updated to include the new model. At the same time, the community‑driven "Ox Alpha" model on OpenRouter and OpenCode, nicknamed “牛来”, sparked discussion for its code‑centric, long‑context (≈1 million tokens) agent capabilities that support text, image, and video inputs with three inference modes (low, high, max).

Since the V4 series preview on April 24, DeepSeek has gradually filled the missing pieces: V4‑Flash entered public beta on July 31, V4‑Pro and the official Agent framework were released on August 13, and the Harness repository now has 178 000 GitHub stars.

Before the official visual model was released, the community built an unofficial “mixed‑vision” version by attaching a third‑party visual encoder to V4‑Flash‑0731.

According to DeepSeek, the V4‑Flash‑Vision‑Exp model matches V4‑Flash in pure‑text ability (including Agent, reasoning, and world knowledge) and shows a large jump on visual‑agent benchmarks, bringing its multimodal agent performance close to Claude Opus‑4.8.

The announcement also provides three concrete capability demos:

Example 1: Using the Agent framework to generate a custom PPT for a high‑net‑worth client’s Tibet self‑drive tour, with specific visual style requirements.

Example 2: Re‑imagining the DeepSeek Harness website with a deep‑sea black‑blue glass UI, guided by a detailed prompt.

Example 3: Creating a front‑end mini‑demo featuring a clay‑monster‑style dynamic effect.

Prompt for Example 1: "Help me make a PPT for a Tibet itinerary, covering south and north, one month, self‑drive, high‑end private service, wild, primitive, exploratory vibe, no typical checkpoints, using real photography, with three realistic pricing options, and a powerful, rugged visual style."

Images accompanying the examples illustrate the generated PPT layout, the redesigned website UI, and the animated clay‑monster demo.

The article concludes by encouraging readers to try the model themselves to assess its capabilities.