Qwen3.6 Docker Stack Hits 118 tok/s

The pitch is less about a new model and more about turning a high-end local LLM setup into something repeatable, portable, and fast enough to matter. For anyone trying to self-host a serious coding model on consumer GPUs, this is the kind of infra packaging that saves days of yak-shaving.

• –The real value is the containerization: it bundles the quant, vLLM flags, model caching, and GPU autodetection into one reproducible path
• –118 tok/s on 2x 3090s is a strong practical benchmark, especially since it reports sustained throughput rather than a cherry-picked peak
• –Keeping the model on a host volume means upgrades are cheaper and less annoying, which is exactly what local-LLM operators want
• –Vision support and OpenAI-compatible API make it more than a toy benchmark stack
• –The main caveat is hardware specificity: the nicest numbers depend on dual 3090-class GPUs and the right vLLM nightlies