DeepSeek Finally Gives Its Agents Eyes
DeepSeek shipped V4-Flash-Vision-Exp on August 21 and Hacker News pushed it past 430 points within hours. It's the multimodal variant of V4 Flash — the smaller, faster half of the V4 family — and it's DeepSeek's first real entry into the vision race.
The specs are classic DeepSeek: sparse mixture-of-experts with 13B active parameters out of 284B total, a full 1,048,576-token context window, and pricing that makes you check twice — $0.22 per million input tokens and $0.66 per million output on OpenRouter. DeepSeek says it comes close to Claude Opus 4.8 on multimodal agent benchmarks while matching the base Flash model on text, reasoning and agent tasks.
The word doing the work in that sentence is "agent." This isn't positioned as a chatbot that can look at your vacation photos. Document understanding, chart reading, visual question answering, and multimodal agent workflows that interleave text and images — that's screenshot parsing, GUI operation, PDF wrangling. The stuff computer-use agents spend all day doing, and the stuff that currently forces you to route to an expensive Western model even when the surrounding text work is cheap.
At $0.22 per million input tokens with a million-token window, you can afford to let an agent look at everything, all the time. If the benchmark claims hold up — it's tagged experimental, so let independent numbers land — the cost floor for vision-enabled agents just dropped through the basement. That's been DeepSeek's move every single time, and it keeps working.
Model details: https://api-docs.deepseek.com/ and on OpenRouter: https://openrouter.ai/deepseek/deepseek-v4-flash-vision-exp
← Back to all articles
The specs are classic DeepSeek: sparse mixture-of-experts with 13B active parameters out of 284B total, a full 1,048,576-token context window, and pricing that makes you check twice — $0.22 per million input tokens and $0.66 per million output on OpenRouter. DeepSeek says it comes close to Claude Opus 4.8 on multimodal agent benchmarks while matching the base Flash model on text, reasoning and agent tasks.
The word doing the work in that sentence is "agent." This isn't positioned as a chatbot that can look at your vacation photos. Document understanding, chart reading, visual question answering, and multimodal agent workflows that interleave text and images — that's screenshot parsing, GUI operation, PDF wrangling. The stuff computer-use agents spend all day doing, and the stuff that currently forces you to route to an expensive Western model even when the surrounding text work is cheap.
At $0.22 per million input tokens with a million-token window, you can afford to let an agent look at everything, all the time. If the benchmark claims hold up — it's tagged experimental, so let independent numbers land — the cost floor for vision-enabled agents just dropped through the basement. That's been DeepSeek's move every single time, and it keeps working.
Model details: https://api-docs.deepseek.com/ and on OpenRouter: https://openrouter.ai/deepseek/deepseek-v4-flash-vision-exp
Comments