DeepSeek Harness v0.1.5 Released Compatible with DeepSeek V4.1 Flash Model
The Zhitong Finance App learned that the DeepSeek Harness team released DeepSeek Harness v0.1.5. The new version is combined with DeepSeek V4.1 Flash model training depth, and the model has been specially trained and optimized in different configurations of DeepSeek Harness. At the same time, the new version of DeepSeek Harness provides plug-in authors with more standardized UI extension portals, adds functions such as file upload and sidebar file preview, optimizes UI performance and interaction, and continues to add experimental features that are deeply integrated with the cutting edge of model research.
Major updates:
Adapting to the DeepSeek V4.1 Flash model
The DeepSeek V4.1 Flash model has been specially trained and optimized for DeepSeek Harness, covering standard mode, programmatic tool call (PTC) mode, and minimal mode.
When using the DeepSeek V4.1 Flash model, the new DeepSeek Harness version also supports updating system prompts while retaining the existing KV Cache.
Feedback on the DeepSeek model and Harness can be submitted to the DeepSeek team using the Like and Tap features and the /feedback command in the UI.
More ways to process and preview files
The web interface now supports uploading various types of files, including images and PDFs, and models can be read as needed using the file tool. The right sidebar supports browsing the workspace file tree and can directly preview files in Markdown, HTML, PDF, common code and image formats; the files produced by the model can be viewed in the sidebar, opened through the system's default application, or located in the file manager.
Standardized web UI extension portal for plug-in authors
The web interface has established standardized plug-in expansion entrances on the left and right sides. The plug-in can register a global portal between the “new session” and “workspace” of the left sidebar and link to the central content panel; the sidebar on the right allows the plug-in to register new tab types and content pages, and display preview content or tool interfaces through multiple tabs, columns, and full screen.
At this point, the plug-in has clear access locations on the left and right sides of the main interface, laying the foundation for subsequent plug-in management functions and the addition of more plug-in expansion portals.
Continuous optimization of long-session performance and user experience
The new version optimizes long session loading, recovery, and continuous conversation performance, reduces memory usage, and optimizes storage formats. Conversation process content folding, complete historical cycle navigation, and interactions such as text width, font size, and operation statistics have also been improved.
Sub-agent communication and task control are more flexible
The parent agent and the child agent, which can continue the conversation, support two-way communication, to supplement information and adjust the direction of tasks during execution; users can also queue to send messages, edit or delete pending messages, and interrupt or stop tasks; and the main agent can select the model and reasoning intensity for the child agent.
Experimental feature Agent Teams
Agent Teams allows the primary agent to create multiple team members to split, assign, and track work through shared task lists. Messages can be exchanged between members; the main agent can view progress, wait for results, or interrupt tasks, and finally summarize. On the web, you can view team members and shared tasks, and you can also enter any member's sub-session to continue communicating.
This function is provided as an experimental plug-in and is disabled by default. Developers need to add relevant experimental plug-ins to the profile in the command line, which will cause additional token consumption.