DeepSeek-V4-Flash-Vision-Exp (2026): Official Launch, Pricing, and a Copy-Paste API Quickstart

From teaser to launch: the 72-hour timeline

Editor's note: an earlier version of this brief, published the morning of Aug 21 before the launch, concluded no official model existed — correct at that hour (docs had no entry; the name was gray-release only), wrong by evening. Same-day official launches are exactly why we date-stamp every claim.

What shipped, per official docs

• The model. Experimental vision-understanding model; model='deepseek-v4-flash-vision-exp'. Official framing: text quality "on par with the official DeepSeek-V4-Flash," multimodal agent capabilities "close to Opus-4.8." V4-Pro did not get vision — this is a Flash-family capability for now.
• Three ways to send images. Base64 data: URLs inline; a public external image URL (8,192-char cap, 60s download timeout); or the Files API — upload once, reference file_id. All three endpoints work: OpenAI-compatible Chat Completions, Anthropic-compatible Messages (Files API needs the anthropic-beta: files-api-2025-04-14 header), and Responses.
• Image limits. JPEG / PNG / GIF / WebP; 32 MiB per image inline or by URL, 64 MiB via file_id; up to 600 images per request; 48 MiB request body (64 MiB total images, 200 MiB with file_id images); max 8,192 px per side (4,096 px when sending 15+ images).
• The Files API (new, separate launch). purpose=user_data; 64 MiB per file; 25 GiB and 10,000 files per user; retention 1 hour to 30 days or permanent; format detected by content, not extension.
• Harness support, same day. DeepSeek Harness v0.1.1-rc.1 (official account, Aug 21 18:22): "DeepSeek 模型适配器新增 DeepSeek-V4-Flash-Vision-Exp 多模态模型选项,并支持配置原生图片请求" — last week's rc.7/rc.8 image plumbing now has a first-party model to plug into (see our rc.7/rc.8 brief).

deepseek-v4-flash-vision-exp pricing (2026): same as V4-Flash, images capped at 384 tokens

deepseek-v4-flash-vision-exp costs the same as deepseek-v4-flash: $0.22 per 1M input tokens off-peak, $0.44 peak, $0.66/$1.32 output — and every image is billed at a maximum of 384 tokens. Peak hours are UTC 01:00–04:00 and 06:00–10:00 (Beijing 09:00–12:00 and 14:00–18:00); off-peak is half of peak.

Source: official pricing page and Vision token-usage guide, read 2026-08-21.

What an image actually costs: vision-exp vs Western vision models (2026)

A maximum-size image on vision-exp costs at most $0.00017 peak ($0.000084 off-peak) — roughly 1,000 screenshots for under $0.09 off-peak. Western vision-capable models bill images by their own (larger, resolution-dependent) token counts at 1.7×–13.6× higher per-token prices:

Comparison prices are list prices from public pricing pages for Claude Sonnet 4.5 and Grok 4.6 and Google's published Gemini 3.7 Flash intro pricing (our Gemini 3.7 brief); competitor image-token counts vary by resolution and are estimates. The structural point stands regardless: vision-exp combines the cheapest per-token price in its class with the only hard per-image token cap.

How to call deepseek-v4-flash-vision-exp: quickstart (curl + Python)

The API is OpenAI-compatible — the standard vision message format works as-is against https://api.deepseek.com. Image by public URL:

curl https://api.deepseek.com/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $DEEPSEEK_API_KEY" \
  -d '{
    "model": "deepseek-v4-flash-vision-exp",
    "messages": [{
      "role": "user",
      "content": [
        {"type": "text", "text": "Describe the UI bug visible in this screenshot."},
        {"type": "image_url",
         "image_url": {"url": "https://example.com/screenshot.png"}}
      ]
    }]
  }'

Same call in Python with the official OpenAI SDK — point base_url at DeepSeek, inline the image as base64:

from openai import OpenAI

client = OpenAI(api_key="YOUR_KEY", base_url="https://api.deepseek.com")

resp = client.chat.completions.create(
    model="deepseek-v4-flash-vision-exp",
    messages=[{
        "role": "user",
        "content": [
            {"type": "text", "text": "Summarize this dashboard."},
            {"type": "image_url",
             "image_url": {"url": "data:image/png;base64,"}},
        ],
    }],
)
print(resp.choices[0].message.content)

• JavaScript/Node: identical shape with the openai npm package — swap baseURL and the model ID.
• Repeated screenshots (agent loops): upload once via the Files API (purpose="user_data"), then reference the returned file_id — 64 MiB per image, no re-upload each turn. This is the pattern DeepSeek Harness v0.1.1-rc.1 uses natively.
• Anthropic-style stack: the Messages endpoint works too; Files API references need the anthropic-beta: files-api-2025-04-14 header.
• Through a relay or gateway: any OpenAI-compatible proxy that forwards the image_url content blocks passes vision payloads unchanged — verify your gateway forwards multi-part content arrays before blaming the model.

Which DeepSeek model to route: vision-exp vs v4-flash vs v4-pro

Get vision-capable Chinese models through one OpenAI-compatible endpoint

Where this leaves the community adapters (and the confusion)

• The name collision is resolved — in DeepSeek's favor. For two weeks, "V4-Flash vision" meant community checkpoints: WebBrain's DeepSeek-V4-Flash-Vision-NVFP4 (frozen V4-Flash + frozen Kimi-K2.6 MoonViT + a 40.1M-parameter trained projector, Aug 3) and FlyCockpit's 0731-vision (Aug 4). The official model now owns the name — with support, docs, and Flash pricing the adapters can't match.
• The adapters aren't worthless — they're repositioned. If you self-host V4-Flash weights on your own GPUs and want offline/air-gapped vision, they remain the only option; the official weights remained closed — until DeepSeek open-sourced Vision-Exp under MIT on Aug 31, 2026, repositioning these adapters as interim community options rather than the only path.
• The gray-to-launch pattern is worth remembering. This is the second DeepSeek release in a month that showed up in API lists or code before docs updated (V4-Flash-0731 followed the same path). For relay operators and tool builders: watching the model-list endpoint and DSH release notes beats waiting for the changelog.

Primary sources

• DeepSeek API changelog — "DeepSeek-V4-Flash-Vision-Exp Release" entry dated 2026-08-21
• DeepSeek official pricing — vision-exp identical to V4-Flash; peak/off-peak windows; concurrency 2,500
• DeepSeek Vision guide — image input methods, limits, 384-token-per-image rule
• DeepSeek Files API guide — formats, 64 MiB/file, 25 GiB & 10k files per user
• DeepSeek Harness 官方公众号 — v0.1.1-rc.1 新增 vision-exp 模型选项与原生图片请求(2026-08-21)
• 爱范儿 — 突发:DeepSeek 多模态模型发布,鲸鱼终于开「天眼」(2026-08-21,含定价截图)
• X — @NFT_Chen:vision-exp 现身 API 列表灰度截图(2026-08-20)
• TestingCatalog — whale-eyes teaser by DeepSeek multimodal researcher Xiaokang Chen (Aug 19–20)
• LINUX DO — 官宣 DeepSeek-V4-Flash-Vision-Exp 发布(2026-08-21,社区一手讨论与价格表)
• Hugging Face — WebBrain community adapter (context: the pre-launch ecosystem)

FAQ (2026)