---
format: "aidr-story-markdown/v1"
id: "ae15427bbb59741f588be3cfb60bca592189608f2488e6c455fcacceba777cc8"
canonical_url: "https://aidr.today/ae15427b?lang=en"
title: "DeepSeek Releases V4 Flash Vision Exp Weights to Rival Opus 4.8 Agent Performance"
lang: "en"
requested_lang: "en"
available_langs: ["en","vi"]
translation_fallback: null
fallback_fields: []
published_at: "2026-08-31T11:53:38.000Z"
category: "Releases"
topics: ["deepseek","agent","multimodal","open-source","harness"]
source_urls: ["https://huggingnews.com/ai/deepseek-releases-v4-flash-vision-exp-weights-to-rival-opus-48-agent-per-95773a98","https://x.com/sgl_project/status/2094377989362663811","https://x.com/MiaAI_lab/status/2094376264152838487","https://x.com/HuggingPapers/status/2094373438278234179","https://x.com/teortaxesTex/status/2094375909868368213","https://x.com/vllm_project/status/2094711861472350343","https://x.com/Ilya_AI_x/status/2094795361948729656","https://huggingnews.com/ai/deepseek-opens-305b-v4-flash-vision-model-near-opus-48-agent-performance-43aea4d0"]
summary: "The company published weights for its DeepSeek-V4-Flash-Vision-Exp model on Hugging Face and its API platform. This experimental multimodal model matches the text capabilities of DeepSeek-V4-Flash in reasoning and world knowledge and brings multimodal agent performance close to Opus-4.8. The release creates feature parity with AI models from Moonshot and GLM. DeepSeek launched Harness 0.1.1 with out-of-the-box support for the model, and the sgl-project is developing an updated cookbook for implementation."
---

# DeepSeek Releases V4 Flash Vision Exp Weights to Rival Opus 4\.8 Agent Performance

> [Open the canonical story](<https://aidr.today/ae15427b?lang=en>)

**Published:** 2026-08-31T11:53:38.000Z
**Category:** Releases
**Topics:** deepseek, agent, multimodal, open\-source, harness

## Summary

The company published weights for its DeepSeek\-V4\-Flash\-Vision\-Exp model on Hugging Face and its API platform\. This experimental multimodal model matches the text capabilities of DeepSeek\-V4\-Flash in reasoning and world knowledge and brings multimodal agent performance close to Opus\-4\.8\. The release creates feature parity with AI models from Moonshot and GLM\. DeepSeek launched Harness 0\.1\.1 with out\-of\-the\-box support for the model, and the sgl\-project is developing an updated cookbook for implementation\.

## Sources

- [Story source](<https://huggingnews.com/ai/deepseek-releases-v4-flash-vision-exp-weights-to-rival-opus-48-agent-per-95773a98>)
- [Story source](<https://x.com/sgl_project/status/2094377989362663811>)
- [Supporting source](<https://x.com/MiaAI_lab/status/2094376264152838487>)
- [Supporting source](<https://x.com/HuggingPapers/status/2094373438278234179>)
- [Supporting source](<https://x.com/teortaxesTex/status/2094375909868368213>)
- [Story source](<https://x.com/vllm_project/status/2094711861472350343>)
- [Supporting source](<https://x.com/Ilya_AI_x/status/2094795361948729656>)
- [Story source](<https://huggingnews.com/ai/deepseek-opens-305b-v4-flash-vision-model-near-opus-48-agent-performance-43aea4d0>)