Top AI Product

DeepSeek shipped its first V4-series vision model on August 21, and HackerNews put it on the front page with 319 points. This is a multimodal model served through the DeepSeek API: text capabilities match DeepSeek-V4-Flash, but on multimodal agent benchmarks it jumps far past its text-only sibling — DeepSeek says close to Claude Opus 4.8.

The price gap is the story

Sparse MoE, 284B total parameters with only 13B active. That’s how you get a 1,048,576-token context window and 384K max output at $0.22/M input and $0.66/M output — a fraction of what frontier multimodal models charge. Images are capped at 384 tokens each, so screenshot-heavy agent loops stay cheap.

API access

Live now as deepseek-v4-flash-vision-exp, also on OpenRouter. Supports Chat Completions, Messages and Responses, with images via base64, URL, or the new Files API (free uploads, reusable file IDs). Built for document and chart understanding, visual QA, and agents that interleave text and images — think an agent reading dashboards or parsing PDFs mid-task.

It’s labeled experimental. If the “Exp” suffix follows DeepSeek’s usual pattern, a stable V4 vision model is coming.

You Might Also Like

• Deepseek tui Tops Github Trending a Claude Code Clone Wired to Deepseeks api
• Web Speed Kills the Token tax for Claude and Gemini Agents Claims 90 Cost cut
• Deepseek v4 Flash 0731 Hits 82 7 on Terminal Bench Chasing Opus 4 8 at 0 14 m Tokens
• Deepseek v4 Flash on a Single amd Mi300x 284b Params one gpu Zero Nvidia
• Anthropic Multi Agent Turf war Research 3 Claude Agents 1 Codebase Self Replicating Malware