qwen3.6-windows-server
A portable Windows launcher that serves Qwen3.6-27B locally through an OpenAI-compatible API with no WSL, Docker, or telemetry, offering a text-based interface to pick between preset speed, context-length, and GPU configurations.
This project packages a patched vLLM wheel and a terminal launcher into a single portable zip that serves Qwen3.6-27B on Windows without WSL, Docker, Conda, pip, or admin rights. The launcher lets a user pick from several measured configuration presets, trading off decode speed, context length, and whether one or two GPUs are used, with arrow keys, and supports adding, editing, and deleting custom presets from inside the interface.
Every preset bundles a tool-calling fix so the resulting OpenAI-compatible endpoint works with Claude Code, Cline, Cursor, Codex, OpenCode, KiloCode, LM Studio, and other compatible clients. On a single RTX 3090 running an INT4 quantization, reported speeds range from about 53 to 72 tokens per second depending on context length, and the launcher automatically detects the number of available GPUs to fall back to single-GPU presets when needed.
It is the companion launcher for the author's separately maintained patched vLLM wheel for Windows.