DeepSeek-v4-flash-vision-exp

	DeepSeek-v4-flash-vision-exp(api-docs.deepseek.com)
	483 points by dares2573 1 day ago \| 151 comments
	tl;dr: DeepSeek released an experimental vision model, `deepseek-v4-flash-vision-exp`, accessible via OpenAI-compatible Chat Completions/Responses APIs and the Anthropic-compatible `/messages` endpoint. It accepts JPEG, PNG, GIF, and WebP images via base64, external URL, or Files API reference, with limits of 32 MiB per inline image (64 MiB via Files API) and up to 600 images per request. Images are resized to ~800×800 before inference, capping usage at 384 tokens per image regardless of original size.
	HN Discussion: ↑Excitement that DeepSeek finally has vision, filling a long-missing capability gap ↓Vision quality is inadequate—fails basic tests like reading clocks compared to competitors ~The 800×800 resize limit is too low for OCR and full-page document use cases •Questioning the product rationale for keeping the text-only variant now that vision exists •Noting the pivot contradicts DeepSeek's earlier stance against multimodal models