Skip to main content

DeepSeek Harness Gets Big Update: Full Multimodal Support and a V4 Tease

Just seven days after its debut, the open-source intelligent agent tool DeepSeek Harness is already back with a significant update. The headline feature? Full multimodal support, which means the DeepSeek model adapter can now handle native image requests. This has sparked plenty of speculation that DeepSeek's own visual model might be just around the corner.

For developers, this update brings some genuinely practical improvements. Previously, commands like /goal and /plan were text-only. Now, you can attach images along with your text, and the @ reference feature has been expanded to include not just file extensions but also historical conversations. So, in a single query, you can throw in a screenshot, a local file, or even a previous chat snippet. To keep complex workflows stable, the new version ensures that image attachments persist across both MCP and ACP, so they don't vanish after a single call. It also fixes issues with oversized images and the payload bloat that came from piling up historical images in multi-turn conversations.

Image

But the update isn't just about multimodal capabilities. There are also notable changes in how components are distributed and how the architecture is optimized. Claude Code and Codex are no longer bundled with the main product; instead, they're now available as Profile Bundles that you can install on demand. Codex, in particular, gets two new tricks: it supports a non-interactive permission mode, which lets it run smoothly in unattended processes, and it supports multiple named instances, so you can run several Codex sub-agents with different configurations and names at the same time. The sub-agent's "report path" has also been refined—reportDelivery can now promptly feedback and wake up the parent task, eliminating the frustrating problem of the parent task blindly waiting for a sub-task to finish.

Image

Other optimizations include concurrent queries for web_search, official support for persistent PowerShell sessions in the Windows PTY terminal (now enabled by default in minimalist mode), and significant improvements to the SQLite backend's read/write and fork performance, along with reduced storage overhead. On top of all that, the official team has confirmed that "DeepSeek Harness" is now a registered trademark, and they've updated the brand usage guidelines accordingly.

This update feels like a clear signal that DeepSeek is doubling down on making Harness a more versatile and robust tool for developers. The multimodal support alone opens up new possibilities for how you interact with the agent, and the architectural tweaks suggest they're listening to user feedback. If the V4 visual model is indeed on the horizon, this update might be the groundwork for something bigger.

Key Points

  • Full Multimodal Support: Commands like /goal and /plan now accept images, and @ references can include historical conversations.
  • Persistent Attachments: Image attachments now persist across MCP and ACP, fixing issues with large images and payload accumulation.
  • Component Distribution: Claude Code and Codex are now optional Profile Bundles, with Codex gaining non-interactive mode and multi-instance support.
  • Performance Boosts: Concurrent web searches, persistent PowerShell sessions, and improved SQLite backend performance.
  • Trademark Registration: "DeepSeek Harness" is now a registered trademark, with updated brand guidelines.