DeepSeek × New Model Releases

DeepSeek launches V4 vision model API

DeepSeek launches V4 vision model API

✎ Story body

DeepSeek shipped a V4 vision API the same week evaluation research shifted toward whether models hold up in practice.

What happened

DeepSeek released the V4-Flash-Vision-Exp multimodal API, with its own docs and a Hacker News thread landing the same day. Of the five articles behind this event, one is official, one community, and three academic — a single vendor post carries the announcement side.

Why it matters

The announcement side is thin; the research side is not. The three papers cover rule grounding for financial compliance, knowledge-grounded domain evaluation, and whether agent memory tracks evolving state — all questions about durability rather than capability. The shift in what gets measured reads as the more durable signal here. Note: no direct link between those papers and DeepSeek is confirmed.

What to watch

Whether the vision API moves from experimental release to GA and real adoption, and whether these evaluation angles show up in vendors' own reported numbers.

▲ Official & Press
Official

DeepSeek-V4-Flash-Vision-Exp Release: Multimodal API Now Live

DeepSeek API Docs / News ・ 2026-08-21 ・ 📌

DeepSeek ships V4-Flash-Vision-Exp, its multimodal API model

Community

DeepSeek-v4-flash-vision-exp

Hacker News (Front Page) ・ 2026-08-21

DeepSeek publishes a Vision guide for its image-capable API

Academic (arxiv etc.) 3 ▾
Academic

OenoBench: A Wine-Domain Benchmark for Knowledge-Grounded Evaluation of Large Language Models

arXiv cs.CL (Computation and Language) ・ 2026-08-20

Academic

ReguSim: Evaluating LLM Agent Rule Grounding in Financial Compliance

arXiv cs.AI (Artificial Intelligence) ・ 2026-08-20

Academic

Can Agent Memory Systems Track Evolving State?

arXiv cs.CL (Computation and Language) ・ 2026-08-20

← Story Archive