Overview
- DeepSeek on Friday released DeepSeek‑V4‑Flash‑Vision‑Exp, an experimental multimodal version of its V4‑Flash line that accepts images and visual prompts through the company’s API.
- The company says internal tests show the model performs close to Anthropic’s Opus 4.8 on multimodal agent tasks, meaning it can reason about images and act with limited human guidance.
- DeepSeek attributes much lower per‑call cost to architectural choices such as a Mixture‑of‑Experts design and a smaller KV cache for image tokens, which it says cuts inference compute versus rival models.
- The lab has published open weights under an MIT license and made the V4 family API‑compatible with OpenAI and Anthropic formats to ease developer switching and speed adoption.
- Industry observers note rising competitive pressure on Western firms and raise questions about training‑from‑output practices, data‑sovereignty risk for customers, and the need for independent third‑party benchmarks.