TODAY

DeepSeek soft-launches vision mode — tester finds knowledge updated past V4 cutoff

DeepSeek quietly added a vision mode to its ChatGPT-style interface, currently rolling out to a fraction of users. Two clues from a hands-on test: the knowledge cutoff extends into April 2026 — newer than the V4 trunk, suggesting an independently trained multimodal model; non-thinking mode responds instantly but hallucinates more, while thinking mode took over four minutes on a spatial reasoning task. OCR and webpage-to-HTML reconstruction both held up well. For readers: a Chinese open-source lab known for text and code reasoning has formally entered the multimodal arena.