DeepSeek Launches V4-Flash-Vision-Exp, Closing In on Opus 4.8
On August 21, DeepSeek launched the experimental multimodal vision model DeepSeek-V4-Flash-Vision-Exp on its official API platform (announced by @deepseekai). Text capabilities match V4-Flash across agents, reasoning and world knowledge, while multimodal agent benchmarks jump substantially—officially described as approaching Anthropic Opus 4.8. Offering near-frontier multimodal agent capability at Flash-tier pricing marks a strategic push into multimodality and agents.
Confirmed
- The model is live on the DeepSeek API platform as an experimental release, with native image input (no image generation) supporting true image understanding rather than OCR only.
- Text capability equals V4-Flash; multimodal agent benchmarks show major gains, approaching Opus 4.8. @卡尔的AI沃茨 also reports improved coding versus the flash version.
- API details: call with model='deepseek-v4-flash-vision-exp', supporting Chat Completions, Messages and Responses interfaces with mixed image-text input; each image costs 117-384 tokens billed at V4-Flash rates.
- Per @petrusenkomax, images can be supplied three ways: inline JPEG/PNG/GIF/WebP (up to 48 MiB), URL (up to 32 MiB), or Files API reference. Per @aigclink's API docs, it supports image description, OCR and chart analysis.
- Per @heypearlai, a Files API shipped simultaneously, free of charge, for uploading images used in conversations; @新智元 and @卡尔的AI沃茨 note the Harness framework was upgraded for image-text input; @新智元 cites a price gap of up to 25x versus competitors (a media claim not in official posts).
- Officials add the model works smoothly across agent frameworks, combining visual understanding with tools for multimodal agent workflows.
- User @teortaxesTex's hands-on testing found reasoning extremely fast; image understanding may not be the strongest, but speed and cost advantages are significant and could disrupt the inference market again.
Why it matters
- A Flash-tier model approaching the frontier closed-source Opus 4.8 on multimodal agent benchmarks extends DeepSeek's cost-performance strategy, directly targeting the low-price, high-speed inference market.
- Cross-framework agent compatibility plus the Harness and free Files API updates let developers plug vision into existing workflows, lowering the barrier for multimodal applications.
2026-08-21 ~ 2026-08-22 · 29 related posts
- Episode 1: DeepSeek Launches V4-Flash-Vision-Exp, Closing In on Opus 4.8(2026-08-21, 29 posts)
- Episode 2: DeepSeek V4 Flash Vision Impresses in Hands-On Tests(2026-08-23, 4 posts)
- Episode 3: DeepSeek Quietly Open-Sources V4-Flash-Vision-Exp, Its First Multimodal V4 Model(2026-08-31, 18 posts)
Primary sources
- DeepSeek Launches V4-Flash Multimodal Model, Agent Capability Near Opus-4.8 — DeepSeek · 2026-08-21
- DeepSeek releases multimodal model with vision agent capabilities rivaling Opus-4.8 — 赛博禅心 · 2026-08-21
- [source] DeepSeek launches V4-Flash-Vision-Exp, multimodal agent performance nears Opus 4.8 — deepseek_ai · 2026-08-21
- DeepSeek: V4-Flash-Vision-Exp works smoothly across agent frameworks — deepseek_ai · 2026-08-21
- [source] DeepSeek Vision API details: up to 384 tokens per image at V4-Flash pricing — deepseek_ai · 2026-08-21
- DeepSeek V4-vision-exp launches API with ultra-fast speed and low cost — teortaxesTex · 2026-08-21
- Deepseek Launches Multimodal Model V4FlashVisionEXP — 歸藏的AI工具箱 · 2026-08-21
- DeepSeek V4 Flash adds vision capabilities, now live — TheCryptoCat75 · 2026-08-21
- DeepSeek launches multimodal API with 384 tokens per image, 25x cheaper than rivals — 新智元 · 2026-08-21
- DeepSeek launches multimodal vision model, rivaling Claude Opus — 机器之心 · 2026-08-21
- DeepSeek Releases V4 Flash Vision API for Image Understanding — aigclink · 2026-08-21
- DeepSeek Adds Vision Support; OpenAI Open Sources Codex Platform — 卡尔的AI沃茨 · 2026-08-21
- DeepSeek releases multimodal model V4-Flash-Vision-Exp with free Files API — heypearlai · 2026-08-21
- DeepSeek Unveils Multimodal Model 'Close to' Anthropic's Opus 4.8 — Polymarket · 2026-08-21
- DeepSeek launches V4 Flash Vision experimental model, performance nears Opus 4.8 — ccerrato147 · 2026-08-21
- DeepSeek opens its eyes: V4-Flash-Vision lands at ~9 images per cent — 量子位 · 2026-08-22
- [source] DeepSeek's v4-flash-vision-exp accepts images inline, via URL, or through Files API — petrusenko_max · 2026-08-22
- DeepSeek V4-Flash-Vision nears Opus-4.8; Anthropic launches $35M security fund — bitdoze · 2026-08-22
11 near-duplicate retellings: teortaxesTex · teortaxesTex · APPSO · 智东西 · eyishazyer · cedric_chee · AI寒武纪 · haider1 · dair_ai · OwariDa · heyshrutimishra