Last night, DeepSeek Harness launched a new version: 14 updates, full of multi-modal capabilities!

Zhidongxi Author | Jiang Yu Editor | Li Shuiqing Zhidongzhi reported on August 20 that last night, DeepSeek Harness received its first important update after the public beta. DeepSeek Harness v0.1.0-rc.8 version is online, and multi-modal capabilities have become the highlight of this update. This update brings a total of 14 adjustments, covering multi-modal input, sub-agent collaboration, terminal experience, tool invocation and developer support. It not only complements the Agent's ability to process multi-modal tasks such as images, but also further improves sub-agent collaboration, terminal interaction and tool invocation experience. The new version supports native image requests and mixed image and text input, and commands such as /goal and /plan can also directly receive images; at the same time, Claude Code and Codex are further connected to its sub-agent system, and a number of issues such as Windows terminal experience, image requests, streaming generation, and custom gateways have also been fixed. DeepSeek Harness officially launched v0.1 version public beta on August 13 and was simultaneously open sourced. On the night of the public beta, Zhixixi immediately pulled the source code for actual testing, and used it to complete tasks such as translation of an 88-page paper and development of the Snake game. In less than a week, this set of Agent Harness has added multi-modality. This update quickly sparked discussion in the DeepSeek Harness community. After seeing the update, some domestic developers sighed directly: "DSH supports multi-modality, please tell me!" Overseas developers also focused on these two capabilities. Some netizens commented that DeepSeek Harness is becoming “more and more interesting”, and multi-modal support and better sub-agent integration are a clear advancement. What’s even more interesting is that some developers have further uncovered the way DeepSeek Harness “views pictures”: for models that do not support image input, it can also call OCR, color statistics, pixel scanning and other tools to split the picture into structured information and then hand it over to the text model for prediction.