Originally published on rohitraj.tech
DeepSeek put image input on its V4-Flash API on August 21, 2026 as deepseek-v4-flash-vision-exp: 284B/13B-active MoE, 1M context, every image capped at 384 tokens at plain V4-Flash rates. Here is the working-developer read — the three ways to send an image, the limits that return 400, the 800×800 downscale that quietly breaks dense OCR, a cost table against Gemini 3.7 Flash and Qwen3.8-Max, and how I would wire it behind a router so the cheap path never becomes the only path.
Read the full version with code samples, diagrams, and architecture details: DeepSeek-V4-Flash-Vision-Exp API Guide: Limits, Pricing, and the 800px Trap (2026)
More engineering notes: rohitraj.tech/en/notes
Top comments (0)