Giving DeepSeek multimodality: the Qwen-3.7-Flash vision approach

Originally posted on X — 89k views, 600+ likes.

Lots of people complain that DeepSeek V4 Flash has no multimodal capability. It's a real gap — but there's no need to be married to one model.

The problem

V4 Flash's text ability is maxed out, but it's helpless with images: screenshots, table images, UI design mockups, scanned documents — none of it works.

Multimodality is the price of "cheap" — the model needs visual encoders, parameters balloon, and costs climb.

The fix: combine models

I surveyed the most cost-effective vision models on the market, and the answer is Qwen-3.7-Flash.

The cost of recognizing one image is negligible. Use it as the dedicated vision model, keep text reasoning on V4 Flash, and get the best of both.

How to combine them

The simplest idea is "routing": input contains an image → vision model; pure text → V4 Flash.

// pseudocode: route on demand
function route(input) {
  if (hasImage(input)) {
    return qwenVision(input)   // recognize image + extract key info
  }
  return deepseekV4Flash(input) // text reasoning
}

A more advanced play is feeding the vision output straight into V4 Flash for further reasoning:

1. Have Qwen-3.7-Flash recognize the image and output a structured description
2. Hand the description + original question to DeepSeek V4 Flash together
3. V4 Flash reasons deeply on the description — writes code, summarizes

This way you get vision and V4 Flash's cheap-and-fast.

Good use cases

• Screenshot Q&A: throw in error screenshots and chat screenshots
• Table / document images into structured data
• UI mockups into code
• Extracting info from scanned invoices and contracts
• Agents that need to "see" a screen

Why not other options

Pure vision LLMs (like the GPT family's multimodal models) are strong, but the price is what it is — not economical for daily batch work.

Qwen-3.7-Flash wins on cost-performance: recognition quality is sufficient, cost is nearly invisible, and you can run big batches without wincing.

Summary

Model combination is the norm — don't expect one model to do everything.

Text workhorse V4 Flash (cheap, fast, huge context), vision supplement Qwen-3.7-Flash (cheap, good enough) — this combo is the best cost-performance option available right now. Patch what's missing rather than waiting for a "does-everything-but-expensive" model to arrive.

Related: DeepSeek V4 Flash — hands-on review of the stable release