Show HN: Swift-Qwen3.8-27B, -58.3% thinking, x1.95 speed, accuracy of xhigh

Swift-Qwen3.8-27B is an optimized 27-billion-parameter Qwen 3.8 variant packaged for low-latency inference, claiming a 58.3% reduction in "thinking" compute, a 1.95× inference speedup, and maintained "xhigh" accuracy. The release targets production-friendly deployments: it presents concrete performance improvements and keeps model behavior control (reasoning effort and accuracy) as first-class options. The name and metrics signal a trade focused on substantially lower runtime cost while preserving high-quality outputs.

The package includes deployment and chat templates (Jinja-based), tokenizer and special-token settings, and configurable system instructions such as reasoning_effort (xhigh/medium/low), tool-interface descriptions, and strict function-call formatting for tool use. Those templates show how to integrate the model into chat-style pipelines and inference endpoints, enforce consistent system/user/assistant roles, and expose options for preserving or suppressing internal chain-of-thought. Practical details - benchmark numbers, inference configuration snippets, and template code - are provided to reproduce the speed and compute claims and to adapt the model for production inference.