DeepSeek begins training models and Agents together: V4.1 Flash specially optimized for Harness

Beating AI News Flash: DeepSeek Harness v0.1.5 has been released. V4.1 Flash underwent specialized training and optimization for Harness during the training phase, covering three modes: Standard, PTC, and Minimalist. Harness has now directly entered the model training process, no longer just an Agent shell attached after model release.

The new version also introduces experimental Agent Teams. Parent and child Agents can communicate bidirectionally, the main Agent can assign different models and reasoning intensities to different child Agents; multiple Agents can also share tasks and send messages to each other, with the main Agent ultimately aggregating the results.

V4.1 Flash also has a very distinctive long-task optimization: if the Agent modifies the system prompt midway through a run, it can still preserve the existing KV Cache as much as possible. Long tasks no longer need to recompute large chunks of subsequent context just because an earlier rule was slightly changed.

The Web interface has also added file upload and preview, supporting PDF, images, code, and other formats. Plugins can integrate their own UI into the left and right Sidebars and the content area, and the loading, restoration, and memory usage of long conversations have also been optimized together.