dsh-model-modes

When using the DeepSeek Harness (DSH) web composer for model inference, developers often face the challenge of flexibly switching Fast mode across multiple AI providers to optimize speed while maintaining fine-grained control over reasoning settings. The dsh-model-modes plugin addresses this by providing capability-aware reasoning profiles and provider-native Fast mode, which decouples Fast mode from reasoning. It intelligently adapts to different provider contracts—such as OpenAI's service_tier or Google's config.serviceTier—applying native fast request parameters without lowering reasoning effort or changing the model, thus simplifying performance optimization in multi-provider environments.

The plugin's core capabilities include automatic expansion of reasoning profiles to all matching providers, support for Max and Ultra as independent reasoning levels, and a clear Fast ON/OFF switch in the composer that only affects the request transmission layer. It also tracks the official model directory to reject unsupported routes and ensures that Fast mode is orthogonal, meaning it never selects cheaper reasoning or alters provider selection. This design enables seamless integration for scenarios requiring efficient inference and precise control, such as real-time applications, batch processing, or cost-sensitive tasks.

This plugin is ideal for developers, AI engineers, and researchers using DeepSeek Harness for model inference, especially those relying on multiple providers like OpenAI, Google, or xAI and needing fast, reliable switching of Fast mode. It streamlines configuration, avoids manual adjustments to models or providers, and allows users to focus on core AI application development, enhancing both workflow efficiency and inference performance.