Skip to main content

DeepSeek Harness Gets Multimodal Upgrade, Hints at V4 Visual Model

It's been only a week since DeepSeek Harness hit the open-source scene, and already it's back with a hefty update. The big news? Full multimodal support. That means the DeepSeek model adapter can now handle native image requests—a move that's got the community whispering about a possible visual model from DeepSeek itself.

But let's talk about what this actually means for developers. Previously, commands like /goal and /plan were strictly text-only. Now, you can toss in images alongside your text. And the @ reference? It's no longer limited to file extensions—you can pull in entire historical conversations. So, in a single query, you could attach a screenshot, a local file, and a chunk of previous chat context. That's a game-changer for debugging or complex tasks.

To keep things stable, the update ensures that image attachments persist across MCP and ACP calls—no more disappearing attachments after a single call. It also fixes issues with oversized images and the payload bloat that came from stacking historical images in multi-turn chats.

Image

But the update isn't just about images. The component distribution got a makeover too. Claude Code and Codex are no longer bundled with the main product. Instead, they're now Profile Bundles you can install on demand. Codex, in particular, got two neat features: non-interactive permission mode, which lets it run unattended, and support for multiple named instances, so you can have several Codex sub-agents with different configs running side by side. Plus, the sub-agent reporting path is smarter now—reportDelivery can ping the parent task promptly, so it's not left waiting blindly for sub-tasks to finish.

Image

Other tweaks? web_search now supports concurrent queries, and the Windows PTY terminal finally has persistent PowerShell sessions—enabled by default in minimalist mode. The SQLite backend got a performance boost for read/write and fork operations, and it takes up less storage. Oh, and the official trademark for "DeepSeek Harness" is now registered, with updated brand usage guidelines.

So, what's the takeaway? This update is a solid step forward for DeepSeek Harness, making it more flexible and efficient. But the real excitement is the hint of a visual model. Could DeepSeek be gearing up to release its own multimodal AI? Only time will tell, but the pieces are falling into place.

Key Points

  • Multimodal support: Native image requests are now possible, hinting at a future visual model.
  • Enhanced commands: /goal and /plan accept images, and @ references can include past conversations.
  • Component changes: Claude Code and Codex are now optional Profile Bundles, with Codex gaining non-interactive mode and multi-instance support.
  • Performance improvements: Concurrent web search, persistent PowerShell sessions, and faster SQLite operations.
  • Trademark registered: "DeepSeek Harness" is now officially trademarked.