DeepSeek Vision API Caps Every Image at 384 Tokens

2 min read
deepseekmultimodalopenai-compatibleapi
View as Markdown
Originally from api-docs.deepseek.com
View source

My notes

Summary

DeepSeek’s deepseek-v4-flash-vision-exp model accepts images via three methods, inline base64, external URL, or a pre-uploaded Files API reference, across OpenAI-compatible, Anthropic-compatible, and Responses API endpoints. Image cost is capped regardless of resolution: every image is resized toward roughly 800x800 pixels before inference, so a 2000x2000 and a 5000x5000 image consume the same fixed 384 tokens.

Key Insight

  • Flat token cost per image is the standout detail. Because resizing targets roughly 800x800 px pre-inference, image cost is capped at 384 tokens no matter the original resolution. Multi-image requests don’t get a discount or penalty; each image is counted independently under the same rule. This makes cost prediction trivial versus providers that scale tokens with resolution or tiling.

  • Three ingestion paths, each with different limits, not just convenience variants:

    • Inline base64: simplest, but counts toward the 48 MiB total request body limit; capped at 32 MiB per image.
    • External URL: the model fetches it; URL max 8192 chars, file max 32 MiB, download must finish in 60 seconds.
    • Files API file_id: the only path that supports images up to 64 MiB and isn’t checked against the 32 MiB per-image cap; best when reusing the same image across many requests (avoids re-upload).

  • Hard ceilings worth knowing before hitting them in production: max 600 images per request; max total image payload 64 MiB without file_id (up to 200 MiB if using file_id references); max image dimension 8192 px per side, but that drops to 4096 px per side once a request contains 15+ images.

  • Format detection is content-based, not filename or MIME-based. A mislabeled file still gets parsed correctly (or rejected) based on actual bytes.

  • Images are restricted to user messages only. Putting an image in a system or assistant message returns a 400 error, and only the vision-tagged model accepts images at all (other models 400 with “This model does not support image”).

  • The detail field (low/high/original/auto) only matters for image_url inputs. low downsamples to 512x512 before inference (cheaper and faster when detail doesn’t matter); it’s silently ignored when the image comes via file_id.

  • The Anthropic-compatible endpoint (/anthropic base path) mirrors the same three methods but reshapes them as image blocks with a source.type of base64, url, or file. The file variant requires the anthropic-beta: files-api-2025-04-14 header.