Glonce

Deepseek releases V4-Flash-Vision-Exp, claims it rivals Opus 4.8 on agent benchmarks

modelsThe Glonce editorsThe Decoder🔥72Glonce rating72Impact68Novelty58Relevance100Surprise48Players76Credibility80
Deepseek releases V4-Flash-Vision-Exp, claims it rivals Opus 4.8 on agent benchmarks

Chinese AI company Deepseek has released Deepseek-V4-Flash-Vision-Exp, an experimental multimodal model that adds image understanding to Deepseek-V4-Flash while keeping the base model's text performance in reasoning and world knowledge, according to Deepseek. On the company's internal multimodal agent benchmarks, the vision variant scores close to Anthropic's Opus 4.8; no named benchmark or numeric score for that comparison is given, only that it nearly matches it.

Deepseek is positioning the model for agent-based visual workflows: it is built to work with different agent frameworks and combine visual understanding with tool use. In practice it can describe images, extract text from screenshots, and analyze diagrams. It handles JPEG, PNG, GIF, and WebP files, and determines the format from the actual file content rather than the filename or declared MIME type, per Deepseek's API docs. The model works with OpenAI's Chat Completions and Responses APIs and Anthropic's Messages endpoint. Alongside it, Deepseek released version 0.1.1 of its Harness framework, which supports the new model out of the box.

There are three ways to send images to the model: embedding them directly with Base64 encoding, pointing to publicly accessible URLs of up to 32 MiB, or using the new, free Files API, which lets a file be uploaded once and referenced by ID across multiple requests, with a size limit of 64 MiB. An optional "detail" field downscales images to 512 x 512 pixels to save tokens when fine visual detail is not needed. Otherwise the model automatically normalizes images to roughly 800 x 800 pixels depending on aspect ratio before processing. Regardless of original resolution, each image costs at most 384 tokens. Pricing follows the existing V4-Flash rates; no dollar figures are given.

A single request can include up to 600 images. The max edge length is 8,192 pixels per side, dropping to 4,096 pixels once a request contains 15 or more images. Images can only appear in user messages.

Key facts

  • Deepseek released Deepseek-V4-Flash-Vision-Exp, an experimental vision model built on V4-Flash that keeps its text reasoning and world-knowledge performance.
  • On Deepseek's internal multimodal agent benchmarks, the vision variant scores close to Anthropic's Opus 4.8; no named benchmark or numeric score is given.
  • The model works with OpenAI's Chat Completions and Responses APIs and Anthropic's Messages endpoint, and Deepseek shipped Harness 0.1.1 to support it.
  • Images can be sent via Base64, public URLs up to 32 MiB, or a new free Files API with a 64 MiB limit; each image costs at most 384 tokens regardless of resolution.
  • A single request can hold up to 600 images, with max edge length dropping from 8,192 to 4,096 pixels once 15 or more images are included.

Why it matters

Deepseek is extending its Flash line from text into multimodal agent work without sacrificing the base model's reasoning and world-knowledge scores, according to the company. Its claim that the vision variant nears Opus 4.8 on internal agent benchmarks signals an attempt to compete directly with Anthropic's frontier models on agentic, tool-using tasks rather than only on raw text benchmarks.

Who it affects

Developers building visual agents, screenshot- or diagram-reading tools, and any team already using OpenAI's or Anthropic's API conventions, since V4-Flash-Vision-Exp is designed to plug into both. It also affects developers using Deepseek's own Harness agent framework, which received a matching 0.1.1 release.

How to use it

The model accepts images through Base64 encoding, public URLs up to 32 MiB, or a new free Files API (64 MiB limit) that lets a file be uploaded once and reused by ID across requests. An optional detail field can downscale images to 512x512 pixels to cut token use; otherwise images are normalized to roughly 800x800 pixels automatically. Each image costs at most 384 tokens no matter its original size, pricing follows existing V4-Flash rates, and a single request can carry up to 600 images with a max edge length of 8,192 pixels (4,096 once 15 or more images are included). Images are only accepted in user messages.

How solid is it

The benchmark claim comes entirely from Deepseek's own internal testing against Opus 4.8, with no named benchmark suite or numeric score disclosed, and no independent verification cited. The model is explicitly labeled experimental ("Exp") by Deepseek itself.

Risks and caveats

The comparison to Opus 4.8 rests on unpublished internal benchmarks chosen and run by Deepseek, which is not the same as a third-party or standardized evaluation. No release date, open-weight status, or dollar pricing is disclosed in the source; pricing is described only as following existing V4-Flash rates.