cutlerbenjamin1-cmd/qwen38-megakernel is a standalone CUDA engine and server that runs Qwen3.8-27B on a single 24GB NVIDIA GPU. The latest release is v2026.10.25. The CUDA megakernel was taken from open-jet, and it also ships a vision encoder and a runner that operate on the CPU. It states that OpenAI-, Anthropic-, and Responses-API-style connections are supported as well. What changed in this release is not described in the given materials.
This project covers an inference method that generates the next token using the model's multi-token prediction head: it proposes and verifies up to 7 draft tokens in a single step. It also covers conversation cache reuse and key/value cache policies, and the README introduces speed and response-matching measurements compared with llama.cpp, along with the testing method. Generation consistency, cache reuse, and matching verification between Windows and Linux are described in separate materials. The license is AGPL-3.0.
That said, since the given materials contain neither the measurements themselves nor a release changelog, it is hard to judge the size of the performance improvement and the significance of this version. It is worth a look as a single-model optimization case in a repository with 3 stars, but the evidence is limited for evaluating it as a general-purpose serving tool.