FULL STORY

DeepSeek's V4-Flash-Vision: From Launch to Open Source

DeepSeek launched the experimental multimodal V4-Flash-Vision-Exp on its API platform on Aug 21, earned strong developer reviews in hands-on tests, and quietly open-sourced it on Hugging Face on Aug 31.

2026-08-21 ~ 2026-09-01 · 3 episodes · 51 posts

Episode 1 · DeepSeek Launches V4-Flash-Vision-Exp, Closing In on Opus 4.8 (2026-08-21, 29 posts)

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.

9 more related posts →

Episode 2 · DeepSeek V4 Flash Vision Impresses in Hands-On Tests (2026-08-23, 4 posts)

After exhausting OpenAI Codex quotas, developers trying DeepSeek-V4-Flash-Vision-Exp reported impressive coding and visual generation results. Comparative tests found Claude Fable 5 superior in layout details for visual generation.

Episode 3 · DeepSeek Quietly Open-Sources V4-Flash-Vision-Exp, Its First Multimodal V4 Model (2026-08-31, 18 posts)

On August 31, DeepSeek quietly released and open-sourced DeepSeek-V4-Flash-Vision-Exp on Hugging Face, the first experimental multimodal model in the V4 series. Built on the V4-Flash architecture with added vision modules and continued training, it is released under MIT with 8-bit and fp8 weights and is simultaneously available via API, marking DeepSeek's entry into multimodal territory.

Confirmed

  • The model was published by @deepseek-ai and is live on Hugging Face (cross-confirmed by @NielsRogge, @victormustar, @ricklamers, @multimodalart, @AdinaYakup and others). It was first spotted by a Reddit user, with no official announcement or documentation at launch (@t4a8945).
  • It is built on the DeepSeek-V4-Flash architecture with vision modules and continued training for visual understanding, making it the first multimodal model in the V4 series (@赛博禅心, @APPSO). Per @solyarisoftware, it adds a 32-layer vision tower and alignment layers on top of the 285B/13B V4-Flash MoE language backbone, with a 1M-token context.
  • Open-sourced under MIT with 8-bit and fp8 variants, Transformers-compatible, in safetensors format (@NielsRogge, @victormustar, @deepseek-ai).
  • Its pipelinetag is image-text-to-text, taking visual and text input for text generation (@AdinaYakup, @victormustar). @zainhas notes it is a sub-500B vision-language experimental model with strong vision-encoding benchmarks on the model card.
  • Per @APPSO, it focuses on Agent vision while retaining text, reasoning and Agent capabilities; per @maxpaperclips, text capability is on par with V4-Flash while multimodal Agent benchmarks approach Opus-4.8; it is usable via API (model='deepseek-v4-flash-vision-exp'). Per @ccerrato147, it scores 83.9 on Terminal Bench 2.1; per @智东西, it wins 5 of 7 text-agent tasks and performs strongly on the DeepSWE software engineering test; per @solyarisoftware, vLLM already supports native inference.

Why it matters

  • This is the first step of the DeepSeek V4 series into multimodality, previously text-only.
  • @teortaxesTex argues the update puts DeepSeek functionally on par with Moonshot (Kimi) and GLM, tightening the multimodal race among leading Chinese open-source models.
  • MIT-licensed weights in multiple precisions lower deployment and fine-tuning barriers, while the simultaneous API launch serves both open and closed paths.