bpx03/qengine is a C++/CUDA inference engine built to run the Qwen3.8-27B INT4 model across multiple RTX 3090 cards. The latest release in the repository is v1.0. The README introduces continuous batching, speculative decoding, KV cache reuse, and image input as supported features. The scope of the project is that these features are provided with that model and hardware configuration.
The main configuration is PCX-based tensor-parallel execution using three RTX 3090s. The server loads a shared library and provides OpenAI- and Anthropic-compatible APIs, and it is described as able to connect to tools such as Claude Code. Performance measurements for a two-card configuration are also published. So rather than a general-purpose inference engine, it is closer to a project that provides an execution path and a way of connecting the server tailored to a specific model and graphics card combination.
However, the published performance measurements presuppose a configuration in which the graphics cards communicate directly with each other through modified NVIDIA kernel drivers. The README states that support for other graphics cards and models is limited. The license is marked as NOASSERTION, so the usage terms cannot be confirmed from this material alone. Rather than applying the performance results as they are to other environments, this is a project best examined within the hardware and driver conditions written in the documentation.