syv-ai/hyperqwen is a repository that provides a vLLM patch and a model preparation pipeline for serving Qwen3.8-27B on a single 24GB GPU. Rather than being limited to one specific deployment configuration, it describes as its components re-quantized weights, a corrected draft vocabulary, and speculative decoding that leverages pieces taken from the prompt. The repository's language is Python and its license is Apache-2.0.
The goal is to run large Qwen models on one consumer GPU. The README presents an OpenAI-compatible API supporting context lengths of 150,000 to 262,000 tokens and key authentication. There are separate Docker Compose configurations for single-user and multi-user setups, and the image size is 9.5GB. It notes that roughly 20GB is needed for model re-quantization on first launch. The patch and scripts to apply and verify it are also included, so the changes can be reverted.
According to the README, on an RTX 3090 at a 250W limit, single-user throughput is 127 tokens per second, and 381 tokens per second when the answer quotes the prompt. With 64 concurrent requests, it reports an aggregate 1,035 tokens per second. These are measurements taken by the author. The document contains no comparison against stock vLLM on the same GPU, so the performance advantage has not been independently confirmed. It is therefore better to treat these numbers as the outcome of a reproducible serving recipe rather than as a verified advantage.