ModelScope(@ModelScope2022):DeepSeek-V4 can see now. 👀 DeepSeek-V4-Flash-Vision-Exp is the family’s first experimental multimodal model. MIT License. 🤖 https://t.co/zbLbIskyzA 🏆 Beats Opus-4.8 on Agents’ Last Exam and ZeroBench, and trails it by just 0.7 on Chartography. 🖼️ Understands images, screenshots, and charts inside tool-using agent workflows. ⚡ Vision added, text strength retained: 83.9 on Terminal-Bench 2.1 and 59.3 on DeepSWE, above Opus-4.8’s 58.0. 🧠 Built on V4-Flash with new visual modules. Weights, prompt encoding, and minimal inference code are open.

ModelScope

@ModelScope2022

Driving innovations with open communities. 💬 Join our Discord:

加入 April 2024

183 正在關注 16K 粉絲

ModelScope@ModelScope2022

2026.08.31 13:46

DeepSeek-V4 can see now. 👀 DeepSeek-V4-Flash-Vision-Exp is the family’s first experimental multimodal model. MIT License. 🤖 🏆 Beats Opus-4.8 on Agents’ Last Exam and ZeroBench, and trails it by just 0.7 on Chartography. 🖼️ Understands images, screenshots, and charts inside tool-using agent workflows. ⚡ Vision added, text strength retained: 83.9 on Terminal-Bench 2.1 and 59.3 on DeepSWE, above Opus-4.8’s 58.0. 🧠 Built on V4-Flash with new visual modules. Weights, prompt encoding, and minimal inference code are open.