DeepSeek Vision Model Nears Opus

DeepSeek is turning its low-cost agent strategy into a multimodal challenge, though the headline comparisons remain vendor-reported and should be independently verified.

• –Supports screenshots, images, and text for visually grounded agent workflows
• –Reports 83.9 on Terminal Bench 2.1, 59.3 on DeepSWE, and 63.6 on DSBench-Hard
• –Visual understanding appears to close a major gap between text-only coding agents and real-world computer use
• –The experimental API release gives developers an inexpensive alternative for UI testing, OCR, and screenshot-driven automation
• –Benchmark caveats matter: several evaluations use DeepSeek’s own Harness and some tests ignore multimodal inputs