Running Qwen 2.5-27B in DeepSeek Harness

Model Integration and Configuration

The Qwen 2.5-27B model functions as a capable local agent when paired with the DeepSeek Harness. To deploy, users configure the harness via a YAML file, specifying the model provider and infrastructure requirements. The author notes that while the model is dense, it achieves usable inference speeds (15-20 tokens per second on single threads, up to 70 tokens per second with concurrency) when deployed on clustered hardware like NVIDIA DGX Spark. The integration allows for full context window utilization and supports speculative decoding using a drafter model.

Vision and Reasoning Capabilities

The model exhibits advanced vision-language processing, specifically in object counting and spatial localization. By dividing images into variable-sized patches, the model identifies and counts objects with high accuracy without requiring external computer vision tools. It can generate Python code to output bounding box coordinates directly. Furthermore, the DeepSeek Harness provides four distinct reasoning effort levels (off, low, medium, extra high). Increasing the reasoning effort significantly improves the quality of complex outputs like website generation, though it causes a non-linear explosion in token consumption. For instance, moving from 'off' to 'medium' reasoning can increase total token usage from approximately 100,000 to nearly 900,000 tokens for the same task.

Observability and Auditing

The DeepSeek Harness includes a built-in trajectory feature that provides full transparency into the agent's decision-making process. This allows developers to audit the initial system prompt, user messages, tool call payloads, and execution timing without needing external tracing tools. This observability is critical for debugging agent failures and refining tool schemas in real-time.