The DeepSeek Harness team has officially rolled out version v0.1.5 of its platform, marking a significant step forward in aligning the tool with the latest DeepSeek V4.1 Flash model.
This new release is deeply integrated with the training pipeline of DeepSeek V4.1 Flash, ensuring the model receives specialized optimization across a range of harness configurations. In addition, the update introduces a host of standardized UI extension points for plugin developers, including file upload capabilities, a sidebar file preview feature, and overall enhancements to interface performance and interactivity. The team has also continued to add experimental features that push the boundaries of frontier model research integration.
Key Updates Tailored for DeepSeek V4.1 Flash
The DeepSeek V4.1 Flash model has undergone dedicated training and tuning specifically for DeepSeek Harness, covering standard mode, programmatic tool call (PTC) mode, and a simplified mode. When operating with DeepSeek V4.1 Flash, this new version of the harness also supports updating system prompts while preserving the existing KV cache. Users can directly submit feedback on both the DeepSeek models and the harness through the UI's like/dislike buttons or by using the /feedback command.
Expanded File Handling and Preview Features
The web interface now allows users to upload a wide array of file types, including images, PDFs, and more. The model can access these files on demand via a dedicated file tool. The right-side sidebar enables browsing of the workspace file tree, with direct previews for Markdown, HTML, PDF, common code files, and image formats. Files generated by the model can be viewed in the sidebar, opened with the system's default application, or located directly in the file manager.
Standardized Plugin Extension Points in Web UI
Standardized plugin extension points have now been established on both the left and right sides of the main web interface. Plugins can register a global entry point on the left sidebar, positioned between the "New Session" and "Workspace" sections, and link it with the central content panel. Meanwhile, the right sidebar supports plugins in registering new tab types and content pages, which can be displayed via multi-tab, split-view, or full-screen layouts for previews or tool interfaces. With this update, plugins have clear integration points on both sides of the main interface, laying the groundwork for future plugin management capabilities and additional extension entry points.
Optimized Performance for Long Conversations
The latest version brings notable improvements to the loading, recovery, and ongoing dialogue performance of long sessions, while also reducing memory usage and refining storage formats. Interactive elements such as content folding, full historical turn navigation, and settings for body width, font size, and runtime statistics have also been polished.
Enhanced Sub-Agent Communication and Task Control
A significant upgrade is the introduction of two-way communication between parent agents and resumable sub-agents. This allows for supplemental information to be provided and task directions to be adjusted mid-execution. Users can also queue messages, edit or delete pending messages, and interject or halt tasks as needed. The main agent can now select both the model and the reasoning intensity for each sub-agent.
Experimental Feature: Agent Teams
An experimental new feature called Agent Teams enables the main agent to create multiple team members who work together via a shared task list, splitting, assigning, and tracking progress. Members can message each other, and the main agent can monitor progress, wait for results, or interrupt tasks before compiling a unified summary. The web interface allows users to view team members and shared tasks, as well as enter any member's sub-conversation to continue the dialogue. This feature is provided as an experimental plugin and is disabled by default; developers must add the relevant plugin to their Profile via the command line to use it. Please note that this feature will incur additional token costs.