DeepSeek Harness v0.1.1 is not just another multimodal endpoint. The framework update adds an adapter for DeepSeek-V4-Flash-Vision-Exp, but the more relevant signal is that images stop being one-shot attachments and become part of a working session's state. The /goal and /plan commands, designed to define objectives and plan steps, now accept text and visual input together. The @ menu can reference files and sessions; MCP and ACP protocols support persistent image attachments; PTC Mode forwards nested images.

This choice has a concrete consequence for teams designing LLM-based pipelines. In a traditional flow, an image is tokenized, processed, and then discarded. With persistent attachments, the framework must serialize the visual reference, manage its lifecycle, and re-present it in subsequent turns without forcing the user to upload the file again. This is not just convenience: it changes how an agent can retain context across documents, diagrams, screenshots, or mockups. For teams working on reporting, code review, or visual analysis, it reduces friction and improves coherence between steps.

The operational trade-off is real. Images consume token budgets much faster than text and, when persistent, affect memory, storage, and inference time. A self-hosted deployment must therefore consider not only the VRAM required by the vision-language model, but also where serialized attachments are stored and how they are retired when the session ends. This is a TCO problem more than a compatibility issue: persistence introduces external state that must be governed, with retention policies and access controls.

There is also a competitive signal. DeepSeek is turning its harness from a simple API-call layer into a runtime for multimodal agentic tasks. The combination of structured commands, file/session references, and persistent attachments brings the tool closer to what demanding enterprise applications need: repeatable workflows where the agent must see, plan, and update a plan. Teams that have already invested in open and self-hosted models can evaluate these mechanisms without relying on cloud APIs, but they must deal with greater infrastructure complexity. For those evaluating on-premise deployment, there are trade-offs between persistence and storage costs that AI-RADAR analyzes in its frameworks at /llm-onpremise.

The direction is clear: multimodality is no longer a chat accessory, but a state component for agents. DeepSeek's 0.1.1 release, although a minor version, shows where the center of gravity is moving: from generating responses to managing persistent visual contexts. The next challenge will not be making images visible to models, but making them survive across turns without exploding costs.