What this is
DeepSeek this week upgraded its open-source tool Harness to v0.1.1. The headline change is vision: it adds a vision-understanding model called DeepSeek-V4-Flash-Vision-Exp (the "Exp" suffix marks it as experimental), with native support for image requests. The /goal and /plan commands can now accept text plus images, and the @ menu can reference files and conversations.
More notably, MCP/ACP protocol integration is now in place. MCP (Model Context Protocol) is the interface standard Anthropic open-sourced last year, letting AI call external tools like Lego bricks; ACP is its evolution. Harness now allows images to stay mounted across MCP sessions, and PTC mode forwards nested images.
In short, DeepSeek is aligning its toolchain with the Agent protocol stack (programs that let AI break down tasks and invoke tools automatically) used by OpenAI and Anthropic.
Industry view
Optimists argue that Chinese LLM companies actively adopting Western-led open standards signals ecosystem maturity—developers won't be locked into any single vendor's models. It also shows DeepSeek is no longer just competing on base models but building out Agent infrastructure.
But there are sober voices too. The version is still 0.1.1—far from stable; the vision model carries an Exp suffix, suggesting a small-scale test. Harness is essentially developer scaffolding, still a long way from enterprise production workflows. Analysts also caution: DeepSeek's updates over the past year have been dense but lack a stunning base-model breakthrough. The honest read is catching up to paths frontier players have already walked—not leading the race.
Impact on regular people
For enterprise IT: Teams already integrated with DeepSeek's API can now layer on image-understanding capabilities—automatically recognizing contract scans or sorting product images, for instance.
For individual professionals: Expect smoother day-to-day use next year—letting AI view screenshots to adjust PPT layouts, read images embedded in English documents, or interpret error screenshots from coworkers.
For consumer markets: End users won't feel it yet, but a maturing open-source toolchain means more startups can build on top of DeepSeek. Consumer AI products' image capabilities will see visible improvement over the next year or two.