AI News Feed
Market watch
LG UltraFine Evo 32U990A-S Hits Record-Low $1,099 on Amazon, TechRadar ReportsNewegg Labor Day Deal: 32GB Kingston Fury Beast DDR4-3200 Kit Nearly Half Price at $179.99 with CodeGalaxy Z Fold 8 inner screen corners reported with excessive giveHuawei unveils Mate XT 2 tri-fold with privacy-screen edition as Apple prepares folding iPhoneWall Street Analysts Recommend Nvidia, Uber and Marvell for Long-Term GrowthLenovo, AMD and Huawei Put Local Large-Model AI at Center of New Devices
LG UltraFine Evo 32U990A-S Hits Record-Low $1,099 on Amazon, TechRadar ReportsNewegg Labor Day Deal: 32GB Kingston Fury Beast DDR4-3200 Kit Nearly Half Price at $179.99 with CodeGalaxy Z Fold 8 inner screen corners reported with excessive giveHuawei unveils Mate XT 2 tri-fold with privacy-screen edition as Apple prepares folding iPhoneWall Street Analysts Recommend Nvidia, Uber and Marvell for Long-Term GrowthLenovo, AMD and Huawei Put Local Large-Model AI at Center of New Devices
Large Language Models

DeepSeek Releases Multimodal Model deepseek-v4-flash-vision-exp, Adds Harness Support

DeepSeek unveils experimental vision model deepseek-v4-flash-vision-exp for agents, with Harness support and Files API.

AI News Feed · 2026-08-21 20:29

The model retains the text capabilities of DeepSeek-V4-Flash, including agent ability, reasoning and world knowledge. On multimodal agent benchmarks it scores significantly higher than V4-Flash, with performance approaching top-end models such as Opus-4.8. It accepts mixed text and image inputs, with images supplied as Base64, external URLs, or through a new Files API.

Developers can call the model through Chat Completions, Messages, and Responses APIs. This opens up tasks such as reading screenshots, analyzing document pages, interpreting table content, identifying interface issues, and generating content from image assets. The pricing follows V4-Flash, with images converted to tokens at a maximum of 384 tokens per image. Input costs range from 0.05 to 3 yuan per million tokens depending on cache and peak hours, and output from 4.5 to 9 yuan; peak hours are 9-12 and 14-18 Beijing time.

The new Files API allows files to be uploaded once and referenced by file_id across requests, avoiding repeated uploads in multi-turn conversations or long-running agent tasks. It supports JPEG, PNG, GIF, and WebP, with a 64 MiB per-file limit, 25 GiB storage per user, and up to 10,000 files.

Earlier the same day, ifanr's guide to DeepSeek Harness showed that the previous multimodal update had prepared the input pipeline but the default DeepSeek-V4-Pro model could not yet read images. The new model completes that chain, enabling Harness to handle image input when using a vision-capable model. The guide also described Harness as a command-line program that can be installed through the npx command or delegated to agents like Codex and WorkBuddy.

DeepSeek positioned the model for multimodal agent scenarios rather than image generation or visual entertainment. The accompanying Files API is aimed at complex agent workflows that need to reference files repeatedly across requests, further expanding the range of information agents can act on.