DeepSeek just dropped an experimental vision model that could complicate things for the big labs. The Chinese company’s new V4-Flash-Vision-Exp takes its existing V4-Flash text model and bolts on image understanding, and the early numbers are eyebrow-raising: on DeepSeek’s internal multimodal agent benchmarks, it scores close to Opus 4.8. That’s a serious flex for a model positioned as a lightweight Flash variant.
The company is aiming squarely at visual agent workflows. The model can describe images, pull text from screenshots, and parse diagrams, and it’s designed to slot into different agent frameworks while combining visual understanding with tool use. Compatibility is broad — it works with OpenAI’s Chat Completions and Responses APIs as well as Anthropic’s Messages endpoint. DeepSeek also shipped version 0.1.1 of its Harness framework with out-of-the-box support for the new model.
Pricing and image handling are where this gets interesting for developers. You can send images three ways: Base64-encoded inline, via publicly accessible URLs up to 32 MiB, or through a new free Files API that lets you upload once and reference by ID across requests, with a 64 MiB cap. A single request can include up to 600 images, and each image costs at most 384 tokens regardless of resolution — the model normalizes everything to roughly 800 x 800 pixels first. There’s also an optional detail field that downscales to 512 x 512 when fine-grained visual information isn’t needed.
What stands out here is the token economics. Capping images at 384 tokens regardless of original resolution is a genuinely useful constraint for agent workloads that process many screenshots or documents. Combined with the 600-image request limit, this looks built for real-world automation pipelines rather than chat demos. The question is whether those internal benchmarks hold up once third parties test it — DeepSeek’s self-reported numbers have been solid before, but the gap between internal evals and independent verification is where credibility gets earned.
